Daniel Rueckert

NeurIPS Conference 2025 Conference Paper

Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks

• Florian Hölzl
• Daniel Rueckert
• Georgios Kaissis

Robust validation metrics remain essential in contemporary deep learning, not only to detect overfitting and poor generalization, but also to monitor training dynamics. In the supervised classification setting, we investigate whether interactions between training data and model weights can yield such a metric that both tracks generalization during training and attributes performance to individual training samples. We introduce Gradient-Weight Alignment (GWA), quantifying the coherence between per-sample gradients and model weights. We show that effective learning corresponds to coherent alignment, while misalignment indicates deteriorating generalization. GWA is efficiently computable during training and reflects both sample-specific contributions and dataset-wide learning dynamics. Extensive experiments show that GWA accurately predicts optimal early stopping, enables principled model comparisons, and identifies influential training samples, providing a validation-set-free approach for model analysis directly from the training data.

TMLR Journal 2025 Journal Article

Improved Localized Machine Unlearning Through the Lens of Memorization

• Reihaneh Torkzadehmahani
• Reza Nasirigerdeh
• Georgios Kaissis
• Daniel Rueckert
• Gintare Karolina Dziugaite
• Eleni Triantafillou

Machine unlearning refers to removing the influence of a specified subset of training data from a model efficiently, after it has already been trained. This is important for key applications, including making the model more accurate by removing outdated, mislabeled, or poisoned data. In this paper, we draw inspiration from prior work that attempts to identify where in the network a given example is memorized, to propose a new "localized unlearning" algorithm, Deletion by Example Localization (DEL). DEL has two components: a localization strategy that identifies critical parameters for a given set of examples, and a simple unlearning algorithm that finetunes only the critical parameters on the data we want to retain. Through extensive experiments, we find that our localization strategy outperforms prior strategies in terms of metrics of interest for unlearning and test accuracy, and pairs well with various unlearning algorithms. Our experiments on different datasets, forget sets, and metrics reveal that DEL outperforms prior work in producing better trade-offs between unlearning performance and accuracy.

ICLR Conference 2025 Conference Paper

Laplace Sample Information: Data Informativeness Through a Bayesian Lens

• Johannes Kaiser
• Kristian Schwethelm
• Daniel Rueckert
• Georgios Kaissis

Accurately estimating the informativeness of individual samples in a dataset is an important objective in deep learning, as it can guide sample selection, which can improve model efficiency and accuracy by removing redundant or potentially harmful samples. We propose $\text{\textit{Laplace Sample Information}}$ ($\mathsf{LSI}$) measure of sample informativeness grounded in information theory widely applicable across model architectures and learning settings. $\mathsf{LSI}$ leverages a Bayesian approximation to the weight posterior and the KL divergence to measure the change in the parameter distribution induced by a sample of interest from the dataset. We experimentally show that $\mathsf{LSI}$ is effective in ordering the data with respect to typicality, detecting mislabeled samples, measuring class-wise informativeness, and assessing dataset difficulty. We demonstrate these capabilities of $\mathsf{LSI}$ on image and text data in supervised and unsupervised settings. Moreover, we show that $\mathsf{LSI}$ can be computed efficiently through probes and transfers well to the training of large models.

ICLR Conference 2025 Conference Paper

SIM: Surface-based fMRI Analysis for Inter-Subject Multimodal Decoding from Movie-Watching Experiments

• Simon Dahan
• Gabriel Bénédict
• Logan Zane John Williams
• Yourong Guo
• Daniel Rueckert
• Robert Leech
• Emma Claire Robinson

Current AI frameworks for brain decoding and encoding, typically train and test models within the same datasets. This limits their utility for cognitive training (neurofeedback) for which it would be useful to pool experiences across individuals to better simulate stimuli not sampled during training. A key obstacle to model generalisation is the degree of variability of inter-subject cortical organisation, which makes it difficult to align or compare cortical signals across participants. In this paper we address this through use of surface vision transformers, which build a generalisable model of cortical functional dynamics, through encoding the topography of cortical networks and their interactions as a moving image across a surface. This is then combined with tri-modal self-supervised contrastive (CLIP) alignment of audio, video, and fMRI modalities to enable the retrieval of visual and auditory stimuli from patterns of cortical activity (and vice-versa). We validate our approach on 7T task-fMRI data from 174 healthy participants engaged in the movie-watching experiment from the Human Connectome Project (HCP). Results show that it is possible to detect which movie clips an individual is watching purely from their brain activity, even for individuals and movies *not seen during training*. Further analysis of attention maps reveals that our model captures individual patterns of brain activity that reflect semantic and visual systems. This opens the door to future personalised simulations of brain function. Code \& pre-trained models will be made available at https://github.com/metrics-lab/sim.

ICLR Conference 2025 Conference Paper

Topograph: An Efficient Graph-Based Framework for Strictly Topology Preserving Image Segmentation

• Laurin Lux
• Alexander H. Berger
• Alexander Weers
• Nico Stucki
• Daniel Rueckert
• Ulrich Bauer
• Johannes C. Paetzold

Topological correctness plays a critical role in many image segmentation tasks, yet most networks are trained using pixel-wise loss functions, such as Dice, neglecting topological accuracy. Existing topology-aware methods often lack robust topological guarantees, are limited to specific use cases, or impose high computational costs. In this work, we propose a novel, graph-based framework for topologically accurate image segmentation that is both computationally efficient and generally applicable. Our method constructs a component graph that fully encodes the topological information of both the prediction and ground truth, allowing us to efficiently identify topologically critical regions and aggregate a loss based on local neighborhood information. Furthermore, we introduce a strict topological metric capturing the homotopy equivalence between the union and intersection of prediction-label pairs. We formally prove the topological guarantees of our approach and empirically validate its effectiveness on binary and multi-class datasets, demonstrating state-of-the-art performance with up to fivefold faster loss computation compared to persistent homology methods.

TMLR Journal 2025 Journal Article

Visual Privacy Auditing with Diffusion Models

• Kristian Schwethelm
• Johannes Kaiser
• Moritz Knolle
• Sarah Lockfisch
• Daniel Rueckert
• Alexander Ziller

Data reconstruction attacks on machine learning models pose a substantial threat to privacy, potentially leaking sensitive information. Although defending against such attacks using differential privacy (DP) provides theoretical guarantees, determining appropriate DP parameters remains challenging. Current formal guarantees on the success of data reconstruction suffer from overly stringent assumptions regarding adversary knowledge about the target data, particularly in the image domain, raising questions about their real-world applicability. In this work, we empirically investigate this discrepancy by introducing a reconstruction attack based on diffusion models (DMs) that only assumes adversary access to real-world image priors and specifically targets the DP defense. We find that (1) real-world data priors significantly influence reconstruction success, (2) current reconstruction bounds do not model the risk posed by data priors well, and (3) DMs can serve as heuristic auditing tools for visualizing privacy leakage.

TMLR Journal 2024 Journal Article

A Survey on Graph Construction for Geometric Deep Learning in Medicine: Methods and Recommendations

• Tamara T. Müller
• Sophie Starck
• Alina Dima
• Stephan Wunderlich
• Kyriaki-Margarita Bintsi
• Kamilia Zaripova
• Rickmer Braren
• Daniel Rueckert

Graph neural networks are powerful tools that enable deep learning on non-Euclidean data structures like graphs, point clouds, and meshes. They leverage the connectivity of data points and can even benefit learning tasks on data, which is not naturally graph-structured -like point clouds. In these cases, the graph structure needs to be determined from the dataset, which adds a significant challenge to the learning process. This opens up a multitude of design choices for creating suitable graph structures, which have a substantial impact on the success of the graph learning task. However, so far no concrete guidance for choosing the most appropriate graph construction is available, not only due to the large variety of methods out there but also because of its strong connection to the dataset at hand. In medicine, for example, a large variety of different data types complicates the selection of graph construction methods even more. We therefore summarise the current state-of-the-art graph construction methods, especially for medical data. In this work, we introduce a categorisation scheme for graph types and graph construction methods. We identify two main strands of graph construction: static and adaptive methods, discuss their advantages and disadvantages, and formulate recommendations for choosing a suitable graph construction method. We furthermore discuss how a created graph structure can be assessed and to what degree it supports graph learning. We hope to support medical research with graph deep learning with this work by elucidating the wide variety of graph construction methods.

TMLR Journal 2024 Journal Article

Are Population Graphs Really as Powerful as Believed?

• Tamara T. Müller
• Sophie Starck
• Kyriaki-Margarita Bintsi
• Alexander Ziller
• Rickmer Braren
• Georgios Kaissis
• Daniel Rueckert

Population graphs and their use in combination with graph neural networks (GNNs) have demonstrated promising results for multi-modal medical data integration and improving disease diagnosis and prognosis. Several different methods for constructing these graphs and advanced graph learning techniques have been established to maximise the predictive power of GNNs on population graphs. However, in this work, we raise the question of whether existing methods are really strong enough by showing that simple baseline methods --such as random forests or linear regressions--, perform on par with advanced graph learning models on several population graph datasets for a variety of different clinical applications. We use the commonly used public population graph datasets TADPOLE and ABIDE, a brain age estimation and a cardiac dataset from the UK Biobank, and a real-world in-house COVID dataset. We (a) investigate the impact of different graph construction methods, graph convolutions, and dataset size and complexity on GNN performance and (b) discuss the utility of GNNs for multi-modal data integration in the context of population graphs. Based on our results, we argue towards the need for "better" graph construction methods or innovative applications for population graphs to render them beneficial.

ICML Conference 2024 Conference Paper

Beyond the Calibration Point: Mechanism Comparison in Differential Privacy

• Georgios Kaissis
• Stefan Kolek
• Borja Balle
• Jamie Hayes
• Daniel Rueckert

In differentially private (DP) machine learning, the privacy guarantees of DP mechanisms are often reported and compared on the basis of a single $(\varepsilon, \delta)$-pair. This practice overlooks that DP guarantees can vary substantially even between mechanisms sharing a given $(\varepsilon, \delta)$, and potentially introduces privacy vulnerabilities which can remain undetected. This motivates the need for robust, rigorous methods for comparing DP guarantees in such cases. Here, we introduce the $\Delta$-divergence between mechanisms which quantifies the worst-case excess privacy vulnerability of choosing one mechanism over another in terms of $(\varepsilon, \delta)$, $f$-DP and in terms of a newly presented Bayesian interpretation. Moreover, as a generalisation of the Blackwell theorem, it is endowed with strong decision-theoretic foundations. Through application examples, we show that our techniques can facilitate informed decision-making and reveal gaps in the current understanding of privacy risks, as current practices in DP-SGD often result in choosing mechanisms with high excess privacy vulnerabilities.

TMLR Journal 2024 Journal Article

Kernel Normalized Convolutional Networks

• Reza Nasirigerdeh
• Reihaneh Torkzadehmahani
• Daniel Rueckert
• Georgios Kaissis

Existing convolutional neural network architectures frequently rely upon batch normalization (BatchNorm) to effectively train the model. BatchNorm, however, performs poorly with small batch sizes, and is inapplicable to differential privacy. To address these limitations, we propose the kernel normalization (KernelNorm) and kernel normalized convolutional layers, and incorporate them into kernel normalized convolutional networks (KNConvNets) as the main building blocks. We implement KNConvNets corresponding to the state-of-the-art ResNets while forgoing the BatchNorm layers. Through extensive experiments, we illustrate that KNConvNets achieve higher or competitive performance compared to the BatchNorm counterparts in image classification and semantic segmentation. They also significantly outperform their batch-independent competitors including those based on layer and group normalization in non-private and differentially private training. Given that, KernelNorm combines the batch-independence property of layer and group normalization with the performance advantage of BatchNorm.

YNICL Journal 2024 Journal Article

LST-AI: A deep learning ensemble for accurate MS lesion segmentation

• Tun Wiltgen
• Julian McGinnis
• Sarah Schlaeger
• Florian Kofler
• CuiCi Voon
• Achim Berthele
• Daria Bischl
• Lioba Grundl

Automated segmentation of brain white matter lesions is crucial for both clinical assessment and scientific research in multiple sclerosis (MS). Over a decade ago, we introduced an engineered lesion segmentation tool, LST. While recent lesion segmentation approaches have leveraged artificial intelligence (AI), they often remain proprietary and difficult to adopt. As an open-source tool, we present LST-AI, an advanced deep learning-based extension of LST that consists of an ensemble of three 3D U-Nets. LST-AI explicitly addresses the imbalance between white matter (WM) lesions and non-lesioned WM. It employs a composite loss function incorporating binary cross-entropy and Tversky loss to improve segmentation of the highly heterogeneous MS lesions. We train the network ensemble on 491 MS pairs of T1-weighted and FLAIR images, collected in-house from a 3T MRI scanner, and expert neuroradiologists manually segmented the utilized lesion maps for training. LST-AI also includes a lesion location annotation tool, labeling lesions as periventricular, infratentorial, and juxtacortical according to the 2017 McDonald criteria, and, additionally, as subcortical. We conduct evaluations on 103 test cases consisting of publicly available data using the Anima segmentation validation tools and compare LST-AI with several publicly available lesion segmentation models. Our empirical analysis shows that LST-AI achieves superior performance compared to existing methods. Its Dice and F1 scores exceeded 0. 62, outperforming LST, SAMSEG (Sequence Adaptive Multimodal SEGmentation), and the popular nnUNet framework, which all scored below 0. 56. Notably, LST-AI demonstrated exceptional performance on the MSSEG-1 challenge dataset, an international WM lesion segmentation challenge, with a Dice score of 0. 65 and an F1 score of 0. 63—surpassing all other competing models at the time of the challenge. With increasing lesion volume, the lesion detection rate rapidly increased with a detection rate of >75% for lesions with a volume between 10 mm3 and 100 mm3. Given its higher segmentation performance, we recommend that research groups currently using LST transition to LST-AI. To facilitate broad adoption, we are releasing LST-AI as an open-source model, available as a command-line tool, dockerized container, or Python script, enabling diverse applications across multiple platforms.

YNICL Journal 2023 Journal Article

Aberrant claustrum structure in preterm-born neonates: an MRI study

• Antonia Neubauer
• Aurore Menegaux
• Jil Wendt
• Hongwei Bran Li
• Benita Schmitz-Koep
• Tobias Ruzok
• Melissa Thalhammer
• David Schinz

The human claustrum is a gray matter structure in the white matter between insula and striatum. Previous analysis found altered claustrum microstructure in very preterm-born adults associated with lower cognitive performance. As the claustrum development is related to hypoxia-ischemia sensitive transient cell populations being at-risk in premature birth, we hypothesized that claustrum structure is already altered in preterm-born neonates. We studied anatomical and diffusion-weighted MRIs of 83 preterm- and 83 term-born neonates at term-equivalent age. Additionally, claustrum development was analyzed both in a spectrum of 377 term-born neonates and longitudinally in 53 preterm-born subjects. Data was provided by the developing Human Connectome Project. Claustrum development showed increasing volume, increasing fractional anisotropy (FA), and decreasing mean diffusivity (MD) around term both across term- and preterm-born neonates. Relative to term-born ones, preterm-born neonates had (i) increased absolute and relative claustrum volumes, both indicating increased cellular and/or extracellular matter and being in contrast to other subcortical gray matter regions of decreased volumes such as thalamus; (ii) lower claustrum FA and higher claustrum MD, pointing at increased extracellular matrix and impaired axonal integrity; and (iii) aberrant covariance between claustrum FA and MD, respectively, and that of distributed gray matter regions, hinting at relatively altered claustrum microstructure. Results together demonstrate specifically aberrant claustrum structure in preterm-born neonates, suggesting altered claustrum development in prematurity, potentially relevant for later cognitive performance.

TMLR Journal 2023 Journal Article

Label Noise-Robust Learning using a Confidence-Based Sieving Strategy

• Reihaneh Torkzadehmahani
• Reza Nasirigerdeh
• Daniel Rueckert
• Georgios Kaissis

In learning tasks with label noise, improving model robustness against overfitting is a pivotal challenge because the model eventually memorizes labels, including the noisy ones. Identifying the samples with noisy labels and preventing the model from learning them is a promising approach to address this challenge. When training with noisy labels, the per-class confidence scores of the model, represented by the class probabilities, can be reliable criteria for assessing whether the input label is the true label or the corrupted one. In this work, we exploit this observation and propose a novel discriminator metric called confidence error and a sieving strategy called CONFES to differentiate between the clean and noisy samples effectively. We provide theoretical guarantees on the probability of error for our proposed metric. Then, we experimentally illustrate the superior performance of our proposed approach compared to recent studies on various settings, such as synthetic and real-world label noise. Moreover, we show CONFES can be combined with other state-of-the-art approaches, such as Co-teaching and DivideMix to further improve model performance.

NeurIPS Conference 2023 Conference Paper

Optimal privacy guarantees for a relaxed threat model: Addressing sub-optimal adversaries in differentially private machine learning

• Georgios Kaissis
• Alexander Ziller
• Stefan Kolek
• Anneliese Riess
• Daniel Rueckert

Differentially private mechanisms restrict the membership inference capabilities of powerful (optimal) adversaries against machine learning models. Such adversaries are rarely encountered in practice. In this work, we examine a more realistic threat model relaxation, where (sub-optimal) adversaries lack access to the exact model training database, but may possess related or partial data. We then formally characterise and experimentally validate adversarial membership inference capabilities in this setting in terms of hypothesis testing errors. Our work helps users to interpret the privacy properties of sensitive data processing systems under realistic threat model relaxations and choose appropriate noise levels for their use-case.

YNIMG Journal 2022 Journal Article

Predicting age and clinical risk from the neonatal connectome

• Yassine Taoudi-Benchekroun
• Daan Christiaens
• Irina Grigorescu
• Oliver Gale-Grant
• Andreas Schuh
• Maximilian Pietsch
• Andrew Chew
• Nicholas Harper

The development of perinatal brain connectivity underpins motor, cognitive and behavioural abilities in later life. Diffusion MRI allows the characterisation of subtle inter-individual differences in structural brain connectivity. Individual brain connectivity maps (connectomes) are by nature high in dimensionality and complex to interpret. Machine learning methods are a powerful tool to uncover properties of the connectome which are not readily visible and can give us clues as to how and why individual developmental trajectories differ. In this manuscript we used Deep Neural Networks and Random Forests to predict demographic and neurodevelopmental characteristics from neonatal structural connectomes in a large sample of babies (n = 524) from the developing Human Connectome Project. We achieved an accurate prediction of post menstrual age (PMA) at scan in term-born infants (mean absolute error (MAE) = 0.72 weeks, r = 0.83 and p < 0.001). We also achieved good accuracy when predicting gestational age at birth in a cohort of term and preterm babies scanned at term equivalent age (MAE = 2.21 weeks, r = 0.82, p < 0.001). We subsequently used sensitivity analysis to obtain feature relevance from our prediction models, with the most important connections for prediction of PMA and GA found to predominantly involve frontal and temporal regions, thalami, and basal ganglia. From our models of PMA at scan for infants born at term, we computed a brain maturation index (predicted age minus actual age) of individual preterm neonates and found a significant correlation between this index and motor outcome at 18 months corrected age. Our results demonstrate the applicability of machine learning techniques in analyses of the neonatal connectome and suggest that a neural substrate of brain maturation with implications for future neurodevelopment is detectable at term equivalent age from the neonatal connectome.

YNIMG Journal 2021 Journal Article

Preterm birth alters the development of cortical microstructure and morphology at term-equivalent age

• Ralica Dimitrova
• Maximilian Pietsch
• Judit Ciarrusta
• Sean P. Fitzgibbon
• Logan Z.J. Williams
• Daan Christiaens
• Lucilio Cordero-Grande
• Dafnis Batalle

INTRODUCTION: The dynamic nature and complexity of the cellular events that take place during the last trimester of pregnancy make the developing cortex particularly vulnerable to perturbations. Abrupt interruption to normal gestation can lead to significant deviations to many of these processes, resulting in atypical trajectory of cortical maturation in preterm birth survivors. METHODS: We sought to first map typical cortical micro- and macrostructure development using invivo MRI in a large sample of healthy term-born infants scanned after birth (n = 259). Then we offer a comprehensive characterization of the cortical consequences of preterm birth in 76 preterm infants scanned at term-equivalent age (37-44 weeks postmenstrual age). We describe the group-average atypicality, the heterogeneity across individual preterm infants, and relate individual deviations from normative development to age at birth and neurodevelopment at 18 months. RESULTS: In the term-born neonatal brain, we observed heterogeneous and regionally specific associations between age at scan and measures of cortical morphology and microstructure, including rapid surface expansion, greater cortical thickness, lower cortical anisotropy and higher neurite orientation dispersion. By term-equivalent age, preterm infants had on average increased cortical tissue water content and reduced neurite density index in the posterior parts of the cortex, and greater cortical thickness anteriorly compared to term-born infants. While individual preterm infants were more likely to show extreme deviations (over 3.1 standard deviations) from normative cortical maturation compared to term-born infants, these extreme deviations were highly variable and showed very little spatial overlap between individuals. Measures of regional cortical development were associated with age at birth, but not with neurodevelopment at 18 months. CONCLUSION: We showed that preterm birth alters cortical micro- and macrostructural maturation near the time of full-term birth. Deviations from normative development were highly variable between individual preterm infants.

YNICL Journal 2020 Journal Article

Parental age effects on neonatal white matter development

• Oliver Gale-Grant
• Daan Christiaens
• Lucilio Cordero-Grande
• Andrew Chew
• Shona Falconer
• Antonios Makropoulos
• Nicholas Harper
• Anthony N Price

OBJECTIVE: Advanced paternal age is associated with poor offspring developmental outcome. Though an increase in paternal age-related germline mutations may affect offspring white matter development, outcome differences could also be due to psychosocial factors. Here we investigate possible cerebral changes prior to strong environmental influences using brain MRI in a cohort of healthy term-born neonates. METHODS: We used structural and diffusion MRI images acquired soon after birth from a cohort (n = 275) of healthy term-born neonates. Images were analysed using a customised tract based spatial statistics (TBSS) processing pipeline. Neurodevelopmental assessment using the Bayley-III scales was offered to all participants at age 18 months. For statistical analysis neonates were compared in two groups, representing the upper quartile (paternal age ≥38 years) and lower three quartiles. The same method was used to assess associations with maternal age. RESULTS: In infants with older fathers (≥38 years), fractional anisotropy, a marker of white matter organisation, was significantly reduced in three early maturing anatomical locations (the corticospinal tract, the corpus callosum, and the optic radiation). Fractional anisotropy in these locations correlated positively with Bayley-III cognitive composite score at 18 months in the advanced paternal age group. A small but significant reduction in total brain volume was also observed in in the infants of older fathers. No significant associations were found between advanced maternal age and neonatal imaging. CONCLUSIONS: The epidemiological association between advanced paternal age and offspring outcome is extremely robust. We have for the first time demonstrated a neuroimaging phenotype of advanced paternal age before sustained parental interaction that correlates with later outcome.

YNICL Journal 2020 Journal Article

Reduced structural connectivity in cortico-striatal-thalamic network in neonates with congenital heart disease

• Megan Ní Bhroin
• Samy Abo Seada
• Alexandra F. Bonthrone
• Christopher J. Kelly
• Daan Christiaens
• Andreas Schuh
• Maximilian Pietsch
• Jana Hutter

Impaired brain development has been observed in newborns with congenital heart disease (CHD). We performed graph theoretical analyses and network-based statistics (NBS) to assess global brain network topology and identify subnetworks of altered connectivity in infants with CHD prior to cardiac surgery. Fifty-eight infants with critical/serious CHD prior to surgery and 116 matched healthy controls as part of the developing Human Connectome Project (dHCP) underwent MRI on a 3T system and high angular resolution diffusion MRI (HARDI) was obtained. Multi-tissue constrained spherical deconvolution, anatomically constrained probabilistic tractography (ACT) and spherical-deconvolution informed filtering of tractograms (SIFT2) was used to construct weighted structural networks. Network topology was assessed and NBS was used to identify structural connectivity differences between CHD and control groups. Structural networks were partitioned into core and peripheral nodes, and edges classed as core, peripheral, or feeder. NBS identified one subnetwork with reduced structural connectivity in CHD infants involving basal ganglia, amygdala, hippocampus, cerebellum, vermis, and temporal and parieto-occipital lobe, primarily affecting core nodes and edges. However, we did not find significantly different global network characteristics in CHD neonates. This locally affected sub-network with reduced connectivity could explain, at least in part, the neurodevelopmental impairments associated with CHD.

YNIMG Journal 2020 Journal Article

The developing Human Connectome Project (dHCP) automated resting-state functional processing framework for newborn infants

• Sean P. Fitzgibbon
• Samuel J. Harrison
• Mark Jenkinson
• Luke Baxter
• Emma C. Robinson
• Matteo Bastiani
• Jelena Bozek
• Vyacheslav Karolis

The developing Human Connectome Project (dHCP) aims to create a detailed 4-dimensional connectome of early life spanning 20–45 weeks post-menstrual age. This is being achieved through the acquisition of multi-modal MRI data from over 1000 in- and ex-utero subjects combined with the development of optimised pre-processing pipelines. In this paper we present an automated and robust pipeline to minimally pre-process highly confounded neonatal resting-state fMRI data, robustly, with low failure rates and high quality-assurance. The pipeline has been designed to specifically address the challenges that neonatal data presents including low and variable contrast and high levels of head motion. We provide a detailed description and evaluation of the pipeline which includes integrated slice-to-volume motion correction and dynamic susceptibility distortion correction, a robust multimodal registration approach, bespoke ICA-based denoising, and an automated QC framework. We assess these components on a large cohort of dHCP subjects and demonstrate that processing refinements integrated into the pipeline provide substantial reduction in movement related distortions, resulting in significant improvements in SNR, and detection of high quality RSNs from neonates.

YNIMG Journal 2019 Journal Article

Automated processing pipeline for neonatal diffusion MRI in the developing Human Connectome Project

• Matteo Bastiani
• Jesper L.R. Andersson
• Lucilio Cordero-Grande
• Maria Murgasova
• Jana Hutter
• Anthony N. Price
• Antonios Makropoulos
• Sean P. Fitzgibbon

The developing Human Connectome Project is set to create and make available to the scientific community a 4-dimensional map of functional and structural cerebral connectivity from 20 to 44 weeks post-menstrual age, to allow exploration of the genetic and environmental influences on brain development, and the relation between connectivity and neurocognitive function. A large set of multi-modal MRI data from fetuses and newborn infants is currently being acquired, along with genetic, clinical and developmental information. In this overview, we describe the neonatal diffusion MRI (dMRI) image processing pipeline and the structural connectivity aspect of the project. Neonatal dMRI data poses specific challenges, and standard analysis techniques used for adult data are not directly applicable. We have developed a processing pipeline that deals directly with neonatal-specific issues, such as severe motion and motion-related artefacts, small brain sizes, high brain water content and reduced anisotropy. This pipeline allows automated analysis of in-vivo dMRI data, probes tissue microstructure, reconstructs a number of major white matter tracts, and includes an automated quality control framework that identifies processing issues or inconsistencies. We here describe the pipeline and present an exemplar analysis of data from 140 infants imaged at 38–44 weeks post-menstrual age.

YNIMG Journal 2018 Journal Article

A review on automatic fetal and neonatal brain MRI segmentation

• Antonios Makropoulos
• Serena J. Counsell
• Daniel Rueckert

In recent years, a variety of segmentation methods have been proposed for automatic delineation of the fetal and neonatal brain MRI. These methods aim to define regions of interest of different granularity: brain, tissue types or more localised structures. Different methodologies have been applied for this segmentation task and can be classified into unsupervised, parametric, classification, atlas fusion and deformable models. Brain atlases are commonly utilised as training data in the segmentation process. Challenges relating to the image acquisition, the rapid brain development as well as the limited availability of imaging data however hinder this segmentation task. In this paper, we review methods adopted for the perinatal brain and categorise them according to the target population, structures segmented and methodology. We outline different methods proposed in the literature and discuss their major contributions. Different approaches for the evaluation of the segmentation accuracy and benchmarks used for the segmentation quality are presented. We conclude this review with a discussion on shortcomings in the perinatal domain and possible future directions.

YNIMG Journal 2018 Journal Article

Construction of a neonatal cortical surface atlas using Multimodal Surface Matching in the Developing Human Connectome Project

• Jelena Bozek
• Antonios Makropoulos
• Andreas Schuh
• Sean Fitzgibbon
• Robert Wright
• Matthew F. Glasser
• Timothy S. Coalson
• Jonathan O'Muircheartaigh

We propose a method for constructing a spatio-temporal cortical surface atlas of neonatal brains aged between 36 and 44 weeks of post-menstrual age (PMA) at the time of scan. The data were acquired as part of the Developing Human Connectome Project (dHCP), and the constructed surface atlases are publicly available. The method is based on a spherical registration approach: Multimodal Surface Matching (MSM), using cortical folding for driving the alignment. Templates have been generated for the anatomical cortical surface and for the cortical feature maps: sulcal depth, curvature, thickness, T1w/T2w myelin maps and cortical regions. To achieve this, cortical surfaces from 270 infants were first projected onto the sphere. Templates were then generated in two stages: first, a reference space was initialised via affine alignment to a group average adult template. Following this, templates were iteratively refined through repeated alignment of individuals to the template space until the variability of the average feature sets converged. Finally, bias towards the adult reference was removed by applying the inverse of the average affine transformations on the template and de-drifting the template. We used temporal adaptive kernel regression to produce age-dependant atlases for 9 weeks (36–44 weeks PMA). The generated templates capture expected patterns of cortical development including an increase in gyrification as well as an increase in thickness and T1w/T2w myelination with increasing age.

YNIMG Journal 2018 Journal Article

Human brain mapping: A systematic comparison of parcellation methods for the human cerebral cortex

• Salim Arslan
• Sofia Ira Ktena
• Antonios Makropoulos
• Emma C. Robinson
• Daniel Rueckert
• Sarah Parisot

The macro-connectome elucidates the pathways through which brain regions are structurally connected or functionally coupled to perform a specific cognitive task. It embodies the notion of representing and understanding all connections within the brain as a network, while the subdivision of the brain into interacting functional units is inherent in its architecture. As a result, the definition of network nodes is one of the most critical steps in connectivity network analysis. Although brain atlases obtained from cytoarchitecture or anatomy have long been used for this task, connectivity-driven methods have arisen only recently, aiming to delineate more homogeneous and functionally coherent regions. This study provides a systematic comparison between anatomical, connectivity-driven and random parcellation methods proposed in the thriving field of brain parcellation. Using resting-state functional MRI data from the Human Connectome Project and a plethora of quantitative evaluation techniques investigated in the literature, we evaluate 10 subject-level and 24 groupwise parcellation methods at different resolutions. We assess the accuracy of parcellations from four different aspects: (1) reproducibility across different acquisitions and groups, (2) fidelity to the underlying connectivity data, (3) agreement with fMRI task activation, myelin maps, and cytoarchitectural areas, and (4) network analysis. This extensive evaluation of different parcellations generated at the subject and group level highlights the strengths and shortcomings of the various methods and aims to provide a guideline for the choice of parcellation technique and resolution according to the task at hand. The results obtained in this study suggest that there is no optimal method able to address all the challenges faced in this endeavour simultaneously.

YNIMG Journal 2018 Journal Article

Metric learning with spectral graph convolutions on brain connectivity networks

• Sofia Ira Ktena
• Sarah Parisot
• Enzo Ferrante
• Martin Rajchl
• Matthew Lee
• Ben Glocker
• Daniel Rueckert

Graph representations are often used to model structured data at an individual or population level and have numerous applications in pattern recognition problems. In the field of neuroscience, where such representations are commonly used to model structural or functional connectivity between a set of brain regions, graphs have proven to be of great importance. This is mainly due to the capability of revealing patterns related to brain development and disease, which were previously unknown. Evaluating similarity between these brain connectivity networks in a manner that accounts for the graph structure and is tailored for a particular application is, however, non-trivial. Most existing methods fail to accommodate the graph structure, discarding information that could be beneficial for further classification or regression analyses based on these similarities. We propose to learn a graph similarity metric using a siamese graph convolutional neural network (s-GCN) in a supervised setting. The proposed framework takes into consideration the graph structure for the evaluation of similarity between a pair of graphs, by employing spectral graph convolutions that allow the generalisation of traditional convolutions to irregular graphs and operates in the graph spectral domain. We apply the proposed model on two datasets: the challenging ABIDE database, which comprises functional MRI data of 403 patients with autism spectrum disorder (ASD) and 468 healthy controls aggregated from multiple acquisition sites, and a set of 2500 subjects from UK Biobank. We demonstrate the performance of the method for the tasks of classification between matching and non-matching graphs, as well as individual subject classification and manifold learning, showing that it leads to significantly improved results compared to traditional methods.

YNIMG Journal 2018 Journal Article

Multimodal surface matching with higher-order smoothness constraints

• Emma C. Robinson
• Kara Garcia
• Matthew F. Glasser
• Zhengdao Chen
• Timothy S. Coalson
• Antonios Makropoulos
• Jelena Bozek
• Robert Wright

In brain imaging, accurate alignment of cortical surfaces is fundamental to the statistical sensitivity and spatial localisation of group studies, and cortical surface-based alignment has generally been accepted to be superior to volume-based approaches at aligning cortical areas. However, human subjects have considerable variation in cortical folding, and in the location of functional areas relative to these folds. This makes alignment of cortical areas a challenging problem. The Multimodal Surface Matching (MSM) tool is a flexible, spherical registration approach that enables accurate registration of surfaces based on a variety of different features. Using MSM, we have previously shown that driving cross-subject surface alignment, using areal features, such as resting state-networks and myelin maps, improves group task fMRI statistics and map sharpness. However, the initial implementation of MSM's regularisation function did not penalize all forms of surface distortion evenly. In some cases, this allowed peak distortions to exceed neurobiologically plausible limits, unless regularisation strength was increased to a level which prevented the algorithm from fully maximizing surface alignment. Here we propose and implement a new regularisation penalty, derived from physically relevant equations of strain (deformation) energy, and demonstrate that its use leads to improved and more robust alignment of multimodal imaging data. In addition, since spherical warps incorporate projection distortions that are unavoidable when mapping from a convoluted cortical surface to the sphere, we also propose constraints that enforce smooth deformation of cortical anatomies. We test the impact of this approach for longitudinal modelling of cortical development for neonates (born between 31 and 43 weeks of post-menstrual age) and demonstrate that the proposed method increases the biological interpretability of the distortion fields and improves the statistical significance of population-based analysis relative to other spherical methods.

ICML Conference 2018 Conference Paper

Semi-Supervised Learning via Compact Latent Space Clustering

• Konstantinos Kamnitsas
• Daniel Coelho de Castro
• Loïc Le Folgoc
• Ian Walker
• Ryutaro Tanno
• Daniel Rueckert
• Ben Glocker
• Antonio Criminisi

We present a novel cost function for semi-supervised learning of neural networks that encourages compact clustering of the latent space to facilitate separation. The key idea is to dynamically create a graph over embeddings of labeled and unlabeled samples of a training batch to capture underlying structure in feature space, and use label propagation to estimate its high and low density regions. We then devise a cost function based on Markov chains on the graph that regularizes the latent space to form a single compact cluster per class, while avoiding to disturb existing clusters during optimization. We evaluate our approach on three benchmarks and compare to state-of-the art with promising results. Our approach combines the benefits of graph-based regularization with efficient, inductive inference, does not require modifications to a network architecture, and can thus be easily applied to existing networks to enable an effective use of unlabeled data.

JBHI Journal 2018 Journal Article

Statistical Shape Modeling of the Left Ventricle: Myocardial Infarct Classification Challenge

• Avan Suinesiaputra
• Jan Dhooge
• Nicolas Duchateau
• Jan Ehrhardt
• Alejandro F. Frangi
• Ali Gooya
• Vicente Grau
• Karim Lekadir

Statistical shape modeling is a powerful tool for visualizing and quantifying geometric and functional patterns of the heart. After myocardial infarction (MI), the left ventricle typically remodels in response to physiological challenges. Several methods have been proposed in the literature to describe statistical shape changes. Which method best characterizes the left ventricular remodeling after MI is an open research question. A better descriptor of remodeling is expected to provide a more accurate evaluation of disease status in MI patients. We therefore designed a challenge to test shape characterization in MI given a set of three-dimensional left ventricular surface points. The training set comprised 100 MI patients, and 100 asymptomatic volunteers (AV). The challenge was initiated in 2015 at the Statistical Atlases and Computational Models of the Heart workshop, in conjunction with the MICCAI conference. The training set with labels was provided to participants, who were asked to submit the likelihood of MI from a different (validation) set of 200 cases (100 AV and 100 MI). Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve were used as the outcome measures. The goals of this challenge were to 1) establish a common dataset for evaluating statistical shape modeling algorithms in MI, and 2) test whether statistical shape modeling provides additional information characterizing MI patients over standard clinical measures. Eleven groups with a wide variety of classification and feature extraction approaches participated in this challenge. All methods achieved excellent classification results with accuracy ranges from 0. 83 to 0. 98. The areas under the receiver operating characteristic curves were all above 0. 90. Four methods showed significantly higher performance than standard clinical measures. The dataset and software for evaluation are available from the Cardiac Atlas Project website. 1 1 http://www.cardiacatlas.org.

YNIMG Journal 2018 Journal Article

The developing human connectome project: A minimal processing pipeline for neonatal cortical surface reconstruction

• Antonios Makropoulos
• Emma C. Robinson
• Andreas Schuh
• Robert Wright
• Sean Fitzgibbon
• Jelena Bozek
• Serena J. Counsell
• Johannes Steinweg

The Developing Human Connectome Project (dHCP) seeks to create the first 4-dimensional connectome of early life. Understanding this connectome in detail may provide insights into normal as well as abnormal patterns of brain development. Following established best practices adopted by the WU-MINN Human Connectome Project (HCP), and pioneered by FreeSurfer, the project utilises cortical surface-based processing pipelines. In this paper, we propose a fully automated processing pipeline for the structural Magnetic Resonance Imaging (MRI) of the developing neonatal brain. This proposed pipeline consists of a refined framework for cortical and sub-cortical volume segmentation, cortical surface extraction, and cortical surface inflation, which has been specifically designed to address considerable differences between adult and neonatal brains, as imaged using MRI. Using the proposed pipeline our results demonstrate that images collected from 465 subjects ranging from 28 to 45 weeks post-menstrual age (PMA) can be processed fully automatically; generating cortical surface models that are topologically correct, and correspond well with manual evaluations of tissue boundaries in 85% of cases. Results improve on state-of-the-art neonatal tissue segmentation models and significant errors were found in only 2% of cases, where these corresponded to subjects with high motion. Downstream, these surfaces will enhance comparisons of functional and diffusion MRI datasets, supporting the modelling of emerging patterns of brain connectivity.

YNIMG Journal 2017 Journal Article

A flexible graphical model for multi-modal parcellation of the cortex

• Sarah Parisot
• Ben Glocker
• Sofia Ira Ktena
• Salim Arslan
• Markus D. Schirmer
• Daniel Rueckert

Advances in neuroimaging have provided a tremendous amount of in-vivo information on the brain's organisation. Its anatomy and cortical organisation can be investigated from the point of view of several imaging modalities, many of which have been studied for mapping functionally specialised cortical areas. There is strong evidence that a single modality is not sufficient to fully identify the brain's cortical organisation. Combining multiple modalities in the same parcellation task has the potential to provide more accurate and robust subdivisions of the cortex. Nonetheless, existing brain parcellation methods are typically developed and tested on single modalities using a specific type of information. In this paper, we propose Graph-based Multi-modal Parcellation (GraMPa), an iterative framework designed to handle the large variety of available input modalities to tackle the multi-modal parcellation task. At each iteration, we compute a set of parcellations from different modalities and fuse them based on their local reliabilities. The fused parcellation is used to initialise the next iteration, forcing the parcellations to converge towards a set of mutually informed modality specific parcellations, where correspondences are established. We explore two different multi-modal configurations for group-wise parcellation using resting-state fMRI, diffusion MRI tractography, myelin maps and task fMRI. Quantitative and qualitative results on the Human Connectome Project database show that integrating multi-modal information yields a stronger agreement with well established atlases and more robust connectivity networks that provide a better representation of the population.

YNICL Journal 2017 Journal Article

Brain lesion segmentation through image synthesis and outlier detection

• Christopher Bowles
• Chen Qin
• Ricardo Guerrero
• Roger Gunn
• Alexander Hammers
• David Alexander Dickie
• Maria Valdés Hernández
• Joanna Wardlaw

Cerebral small vessel disease (SVD) can manifest in a number of ways. Many of these result in hyperintense regions visible on T 2-weighted magnetic resonance (MR) images. The automatic segmentation of these lesions has been the focus of many studies. However, previous methods tended to be limited to certain types of pathology, as a consequence of either restricting the search to the white matter, or by training on an individual pathology. Here we present an unsupervised abnormality detection method which is able to detect abnormally hyperintense regions on FLAIR regardless of the underlying pathology or location. The method uses a combination of image synthesis, Gaussian mixture models and one class support vector machines, and needs only be trained on healthy tissue. We evaluate our method by comparing segmentation results from 127 subjects with SVD with three established methods and report significantly superior performance across a number of metrics.

YNICL Journal 2017 Journal Article

Five-class differential diagnostics of neurodegenerative diseases using random undersampling boosting

• Tong Tong
• Christian Ledig
• Ricardo Guerrero
• Andreas Schuh
• Juha Koikkalainen
• Antti Tolonen
• Hanneke Rhodius
• Frederik Barkhof

Differentiating between different types of neurodegenerative diseases is not only crucial in clinical practice when treatment decisions have to be made, but also has a significant potential for the enrichment of clinical trials. The purpose of this study is to develop a classification framework for distinguishing the four most common neurodegenerative diseases, including Alzheimer's disease, frontotemporal lobe degeneration, Dementia with Lewy bodies and vascular dementia, as well as patients with subjective memory complaints. Different biomarkers including features from images (volume features, region-wise grading features) and non-imaging features (CSF measures) were extracted for each subject. In clinical practice, the prevalence of different dementia types is imbalanced, posing challenges for learning an effective classification model. Therefore, we propose the use of the RUSBoost algorithm in order to train classifiers and to handle the class imbalance training problem. Furthermore, a multi-class feature selection method based on sparsity is integrated into the proposed framework to improve the classification performance. It also provides a way for investigating the importance of different features and regions. Using a dataset of 500 subjects, the proposed framework achieved a high accuracy of 75.2% with a balanced accuracy of 69.3% for the five-class classification using ten-fold cross validation, which is significantly better than the results using support vector machine or random forest, demonstrating the feasibility of the proposed framework to support clinical decision making.

YNICL Journal 2017 Journal Article

Fully automatic acute ischemic lesion segmentation in DWI using convolutional neural networks

• Liang Chen
• Paul Bentley
• Daniel Rueckert

Stroke is an acute cerebral vascular disease, which is likely to cause long-term disabilities and death. Acute ischemic lesions occur in most stroke patients. These lesions are treatable under accurate diagnosis and treatments. Although diffusion-weighted MR imaging (DWI) is sensitive to these lesions, localizing and quantifying them manually is costly and challenging for clinicians. In this paper, we propose a novel framework to automatically segment stroke lesions in DWI. Our framework consists of two convolutional neural networks (CNNs): one is an ensemble of two DeconvNets (Noh et al., 2015), which is the EDD Net; the second CNN is the multi-scale convolutional label evaluation net (MUSCLE Net), which aims to evaluate the lesions detected by the EDD Net in order to remove potential false positives. To the best of our knowledge, it is the first attempt to solve this problem and using both CNNs achieves very good results. Furthermore, we study the network architectures and key configurations in detail to ensure the best performance. It is validated on a large dataset comprising clinical acquired DW images from 741 subjects. A mean accuracy of Dice coefficient obtained is 0.67 in total. The mean Dice scores based on subjects with only small and large lesions are 0.61 and 0.83, respectively. The lesion detection rate achieved is 0.94.

YNIMG Journal 2016 Journal Article

An exploration of task based fMRI in neonates using echo-shifting to allow acquisition at longer T without loss of temporal efficiency

• Giulio Ferrazzi
• Rita G. Nunes
• Tomoki Arichi
• Andreia S. Gaspar
• Giovanni Barone
• Alessandro Allievi
• Serge Vasylechko
• Maryam Abaei

Optimal contrast to noise ratio of the BOLD signal in neonatal and foetal fMRI has been hard to achieve because of the much longer T 2 ⁎ values in developing brain tissue in comparison to those in the mature adult brain. The conventional approach of optimizing fMRI sequences would suggest matching the echo time (T E ) and the T 2 ⁎ of the neonatal and foetal brain. However, the use of a long echo time would typically increase the minimum repetition time (T R ) resulting in inefficient sampling. Here we apply the concept of echo shifting to task based neonatal fMRI in order to achieve an improved contrast to noise ratio and efficient data sampling at the same time. Echo shifted EPI (es-EPI) is a modification of a standard 2D-EPI sequence which enables echo times longer than the time between consecutive excitations ( T E > T S = T R N S, where N S is the number of acquired slices and T S the inter-slice repetition time). The proposed method was tested on neonatal subjects using a passive sensori-motor task paradigm. Dual echo EPI datasets with an identical readout structure to es-EPI were also acquired and used as control data to assess BOLD activation. From the results of the latter analysis, an average increase of 78±41% in contrast to noise ratio was observable when comparing late to short echoes. Furthermore, es-EPI allowed the acquisition of data with an identical contrast to the late echo, but more efficiently since a higher number of slices could be acquired in the same amount of time.

YNICL Journal 2016 Journal Article

Differential diagnosis of neurodegenerative diseases using structural MRI data

• Juha Koikkalainen
• Hanneke Rhodius-Meester
• Antti Tolonen
• Frederik Barkhof
• Betty Tijms
• Afina W. Lemstra
• Tong Tong
• Ricardo Guerrero

Different neurodegenerative diseases can cause memory disorders and other cognitive impairments. The early detection and the stratification of patients according to the underlying disease are essential for an efficient approach to this healthcare challenge. This emphasizes the importance of differential diagnostics. Most studies compare patients and controls, or Alzheimer's disease with one other type of dementia. Such a bilateral comparison does not resemble clinical practice, where a clinician is faced with a number of different possible types of dementia. Here we studied which features in structural magnetic resonance imaging (MRI) scans could best distinguish four types of dementia, Alzheimer's disease, frontotemporal dementia, vascular dementia, and dementia with Lewy bodies, and control subjects. We extracted an extensive set of features quantifying volumetric and morphometric characteristics from T1 images, and vascular characteristics from FLAIR images. Classification was performed using a multi-class classifier based on Disease State Index methodology. The classifier provided continuous probability indices for each disease to support clinical decision making. A dataset of 504 individuals was used for evaluation. The cross-validated classification accuracy was 70.6% and balanced accuracy was 69.1% for the five disease groups using only automatically determined MRI features. Vascular dementia patients could be detected with high sensitivity (96%) using features from FLAIR images. Controls (sensitivity 82%) and Alzheimer's disease patients (sensitivity 74%) could be accurately classified using T1-based features, whereas the most difficult group was the dementia with Lewy bodies (sensitivity 32%). These results were notable better than the classification accuracies obtained with visual MRI ratings (accuracy 44.6%, balanced accuracy 51.6%). Different quantification methods provided complementary information, and consequently, the best results were obtained by utilizing several quantification methods. The results prove that automatic quantification methods and computerized decision support methods are feasible for clinical practice and provide comprehensive information that may help clinicians in the diagnosis making.

YNIMG Journal 2016 Journal Article

Group-wise parcellation of the cortex through multi-scale spectral clustering

• Sarah Parisot
• Salim Arslan
• Jonathan Passerat-Palmbach
• William M. Wells
• Daniel Rueckert

The delineation of functionally and structurally distinct regions as well as their connectivity can provide key knowledge towards understanding the brain's behaviour and function. Cytoarchitecture has long been the gold standard for such parcellation tasks, but has poor scalability and cannot be mapped in vivo. Functional and diffusion magnetic resonance imaging allow in vivo mapping of brain's connectivity and the parcellation of the brain based on local connectivity information. Several methods have been developed for single subject connectivity driven parcellation, but very few have tackled the task of group-wise parcellation, which is essential for uncovering group specific behaviours. In this paper, we propose a group-wise connectivity-driven parcellation method based on spectral clustering that captures local connectivity information at multiple scales and directly enforces correspondences between subjects. The method is applied to diffusion Magnetic Resonance Imaging driven parcellation on two independent groups of 50 subjects from the Human Connectome Project. Promising quantitative and qualitative results in terms of information loss, modality comparisons, group consistency and inter-group similarities demonstrate the potential of the method.

YNIMG Journal 2016 Journal Article

Regional growth and atlasing of the developing human brain

• Antonios Makropoulos
• Paul Aljabar
• Robert Wright
• Britta Hüning
• Nazakat Merchant
• Tomoki Arichi
• Nora Tusor
• Joseph V. Hajnal

Detailed morphometric analysis of the neonatal brain is required to characterise brain development and define neuroimaging biomarkers related to impaired brain growth. Accurate automatic segmentation of neonatal brain MRI is a prerequisite to analyse large datasets. We have previously presented an accurate and robust automatic segmentation technique for parcellating the neonatal brain into multiple cortical and subcortical regions. In this study, we further extend our segmentation method to detect cortical sulci and provide a detailed delineation of the cortical ribbon. These detailed segmentations are used to build a 4-dimensional spatio-temporal structural atlas of the brain for 82 cortical and subcortical structures throughout this developmental period. We employ the algorithm to segment an extensive database of 420 MR images of the developing brain, from 27 to 45weeks post-menstrual age at imaging. Regional volumetric and cortical surface measurements are derived and used to investigate brain growth and development during this critical period and to assess the impact of immaturity at birth. Whole brain volume, the absolute volume of all structures studied, cortical curvature and cortical surface area increased with increasing age at scan. Relative volumes of cortical grey matter, cerebellum and cerebrospinal fluid increased with age at scan, while relative volumes of white matter, ventricles, brainstem and basal ganglia and thalami decreased. Preterm infants at term had smaller whole brain volumes, reduced regional white matter and cortical and subcortical grey matter volumes, and reduced cortical surface area compared with term born controls, while ventricular volume was greater in the preterm group. Increasing prematurity at birth was associated with a reduction in total and regional white matter, cortical and subcortical grey matter volume, an increase in ventricular volume, and reduced cortical surface area.

YNICL Journal 2014 Journal Article

Prediction of stroke thrombolysis outcome using CT brain machine learning

• Paul Bentley
• Jeban Ganesalingam
• Anoma Lalani Carlton Jones
• Kate Mahady
• Sarah Epton
• Paul Rinne
• Pankaj Sharma
• Omid Halse

A critical decision-step in the emergency treatment of ischemic stroke is whether or not to administer thrombolysis — a treatment that can result in good recovery, or deterioration due to symptomatic intracranial haemorrhage (SICH). Certain imaging features based upon early computerized tomography (CT), in combination with clinical variables, have been found to predict SICH, albeit with modest accuracy. In this proof-of-concept study, we determine whether machine learning of CT images can predict which patients receiving tPA will develop SICH as opposed to showing clinical improvement with no haemorrhage. Clinical records and CT brains of 116 acute ischemic stroke patients treated with intravenous thrombolysis were collected retrospectively (including 16 who developed SICH). The sample was split into training (n = 106) and test sets (n = 10), repeatedly for 1760 different combinations. CT brain images acted as inputs into a support vector machine (SVM), along with clinical severity. Performance of the SVM was compared with established prognostication tools (SEDAN and HAT scores; original, or after adaptation to our cohort). Predictive performance, assessed as area under receiver-operating-characteristic curve (AUC), of the SVM (0.744) compared favourably with that of prognostic scores (original and adapted versions: 0.626–0.720; p < 0.01). The SVM also identified 9 out of 16 SICHs, as opposed to 1–5 using prognostic scores, assuming a 10% SICH frequency (p < 0.001). In summary, machine learning methods applied to acute stroke CT images offer automation, and potentially improved performance, for prediction of SICH following thrombolysis. Larger-scale cohorts, and incorporation of advanced imaging, should be tested with such methods.

YNIMG Journal 2013 Journal Article

Random forest-based similarity measures for multi-modal classification of Alzheimer's disease

• Katherine R. Gray
• Paul Aljabar
• Rolf A. Heckemann
• Alexander Hammers
• Daniel Rueckert

Neurodegenerative disorders, such as Alzheimer's disease, are associated with changes in multiple neuroimaging and biological measures. These may provide complementary information for diagnosis and prognosis. We present a multi-modality classification framework in which manifolds are constructed based on pairwise similarity measures derived from random forest classifiers. Similarities from multiple modalities are combined to generate an embedding that simultaneously encodes information about all the available features. Multi-modality classification is then performed using coordinates from this joint embedding. We evaluate the proposed framework by application to neuroimaging and biological data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Features include regional MRI volumes, voxel-based FDG-PET signal intensities, CSF biomarker measures, and categorical genetic information. Classification based on the joint embedding constructed using information from all four modalities out-performs the classification based on any individual modality for comparisons between Alzheimer's disease patients and healthy controls, as well as between mild cognitive impairment patients and healthy controls. Based on the joint embedding, we achieve classification accuracies of 89% between Alzheimer's disease patients and healthy controls, and 75% between mild cognitive impairment patients and healthy controls. These results are comparable with those reported in other recent studies using multi-kernel learning. Random forests provide consistent pairwise similarity measures for multiple modalities, thus facilitating the combination of different types of feature data. We demonstrate this by application to data in which the number of features differs by several orders of magnitude between modalities. Random forest classifiers extend naturally to multi-class problems, and the framework described here could be applied to distinguish between multiple patient groups in the future.

YNIMG Journal 2013 Journal Article

Segmentation of MR images via discriminative dictionary learning and sparse coding: Application to hippocampus labeling

• Tong Tong
• Robin Wolz
• Pierrick Coupé
• Joseph V. Hajnal
• Daniel Rueckert

We propose a novel method for the automatic segmentation of brain MRI images by using discriminative dictionary learning and sparse coding techniques. In the proposed method, dictionaries and classifiers are learned simultaneously from a set of brain atlases, which can then be used for the reconstruction and segmentation of an unseen target image. The proposed segmentation strategy is based on image reconstruction, which is in contrast to most existing atlas-based labeling approaches that rely on comparing image similarities between atlases and target images. In addition, we propose a Fixed Discriminative Dictionary Learning for Segmentation (F-DDLS) strategy, which can learn dictionaries offline and perform segmentations online, enabling a significant speed-up in the segmentation stage. The proposed method has been evaluated for the hippocampus segmentation of 80 healthy ICBM subjects and 202 ADNI images. The robustness of the proposed method, especially of our F-DDLS strategy, was validated by training and testing on different subject groups in the ADNI database. The influence of different parameters was studied and the performance of the proposed method was also compared with that of the nonlocal patch-based approach. The proposed method achieved a median Dice coefficient of 0. 879 on 202 ADNI images and 0. 890 on 80 ICBM subjects, which is competitive compared with state-of-the-art methods.

YNIMG Journal 2012 Journal Article

Automated measurement of local white matter lesion volume

• Fedde van der Lijn
• Benjamin F.J. Verhaaren
• M. Arfan Ikram
• Stefan Klein
• Marleen de Bruijne
• Henri A. Vrooman
• Meike W. Vernooij
• Alexander Hammers

It has been hypothesized that white matter lesions at different locations may have different etiology and clinical consequences. Several approaches for the quantification of local white matter lesion load have been proposed in the literature, most of which rely on a distinction between lesions in a periventricular region close to the ventricles and a subcortical zone further away. In this work we present a novel automated method for local white matter lesion volume quantification in magnetic resonance images. The method segments and measures the white matter lesion volume in 43 regions defined by orientation and distance to the ventricles, which allows a more spatially detailed study of lesion load. The potential of the method was demonstrated by analyzing the effect of blood pressure on the regional white matter lesion volume in 490 elderly subjects taken from a longitudinal population study. The method was also compared to two commonly used techniques to assess the periventricular and subcortical lesion load. The main finding was that high blood pressure was primarily associated with lesion load in the vascular watershed area that forms the border between the periventricular and subcortical regions. It explains the associations found for both the periventricular and subcortical load computed for the same data, and that were reported in the literature. But the proposed method can localize the region of association with greater precision than techniques that distinguish between periventricular and subcortical lesions only.

YNIMG Journal 2012 Journal Article

Construction of a consistent high-definition spatio-temporal atlas of the developing brain using adaptive kernel regression

• Ahmed Serag
• Paul Aljabar
• Gareth Ball
• Serena J. Counsell
• James P. Boardman
• Mary A. Rutherford
• A. David Edwards
• Joseph V. Hajnal

Medical imaging has shown that, during early development, the brain undergoes more changes in size, shape and appearance than at any other time in life. A better understanding of brain development requires a spatio-temporal atlas that characterizes the dynamic changes during this period. In this paper we present an approach for constructing a 4D atlas of the developing brain, between 28 and 44weeks post-menstrual age at time of scan, using T1 and T2 weighted MR images from 204 premature neonates. The method used for the creation of the average 4D atlas utilizes non-rigid registration between all pairs of images to eliminate bias in the atlas toward any of the original images. In addition, kernel regression is used to produce age-dependent anatomical templates. A novelty in our approach is the use of a time-varying kernel width, to overcome the variations in the distribution of subjects at different ages. This leads to an atlas that retains a consistent level of detail at every time-point. Comparisons between the resulting atlas and atlases constructed using affine and non-rigid registration are presented. The resulting 4D atlas has greater anatomic definition than currently available 4D atlases created using various affine and non-rigid registration approaches, an important factor in improving registrations between the atlas and individual subjects. Also, the resulting 4D atlas can serve as a good representative of the population of interest as it reflects both global and local changes. The atlas is publicly available at www. brain-development. org.

YNIMG Journal 2012 Journal Article

Erratum to Construction of a consistent high-definition spatio-temporal atlas of the developing brain using adaptive kernel regression [NeuroImage 59/3(2012) 2255–2265]

• Ahmed Serag
• Paul Aljabar
• Gareth Ball
• Serena J. Counsell
• James P. Boardman
• Mary A. Rutherford
• A. David Edwards
• Joseph V. Hajnal

YNIMG Journal 2012 Journal Article

Magnetic resonance imaging of the newborn brain: Manual segmentation of labelled atlases in term-born and preterm infants

• Ioannis S. Gousias
• A. David Edwards
• Mary A. Rutherford
• Serena J. Counsell
• Jo V. Hajnal
• Daniel Rueckert
• Alexander Hammers

Premature birth is a major and growing problem. Investigations into neuroanatomical correlates and consequences of preterm birth are hampered by complex neonatal brain anatomy and unavailability of atlases and protocols covering the whole brain. We developed delineation protocols for the manual segmentation of cerebral magnetic resonance (MR) images from newborn infants into 50 regions with comprehensive coverage of the brain. We then segmented MR scans from 15 infants born preterm at median 29, range 26–35, weeks postmenstrual age and scanned at term-corrected age, and five term-born infants born at median 41, range 39–45, weeks postmenstrual age. Total and regional brain volumes were estimated in each infant, and regional volumes expressed as a fraction of total brain volume. Total brain volumes were higher with greater age at birth and at time of scan, but once corrected for age at scan there was no difference between preterm and term infants. Fractional age-corrected regional volumes were bigger unilaterally in terms in middle and inferior temporal gyri, anterior temporal lobe, fusiform gyrus and posterior cingulate gyrus. Fractional age-corrected regional volumes were larger in preterms bilaterally in hippocampus, amygdala, thalamus and lateral ventricles, left superior temporal gyrus and right caudate nucleus. These differences were not significant after correcting for multiple hypothesis testing, but suggest subtle differences between preterms and term-borns accessible to regional analysis. Detailed illustrated protocols are made available in the Appendix.

YNIMG Journal 2012 Journal Article

Multi-region analysis of longitudinal FDG-PET for the classification of Alzheimer's disease

• Katherine R. Gray
• Robin Wolz
• Rolf A. Heckemann
• Paul Aljabar
• Alexander Hammers
• Daniel Rueckert

Imaging biomarkers for Alzheimer's disease are desirable for improved diagnosis and monitoring, as well as drug discovery. Automated image-based classification of individual patients could provide valuable diagnostic support for clinicians, when considered alongside cognitive assessment scores. We investigate the value of combining cross-sectional and longitudinal multi-region FDG-PET information for classification, using clinical and imaging data from the Alzheimer's Disease Neuroimaging Initiative. Whole-brain segmentations into 83 anatomically defined regions were automatically generated for baseline and 12-month FDG-PET images. Regional signal intensities were extracted at each timepoint, as well as changes in signal intensity over the follow-up period. Features were provided to a support vector machine classifier. By combining 12-month signal intensities and changes over 12months, we achieve significantly increased classification performance compared with using any of the three feature sets independently. Based on this combined feature set, we report classification accuracies of 88% between patients with Alzheimer's disease and elderly healthy controls, and 65% between patients with stable mild cognitive impairment and those who subsequently progressed to Alzheimer's disease. We demonstrate that information extracted from serial FDG-PET through regional analysis can be used to achieve state-of-the-art classification of diagnostic groups in a realistic multi-centre setting. This finding may be usefully applied in the diagnosis of Alzheimer's disease, predicting disease course in individuals with mild cognitive impairment, and in the selection of participants for clinical trials.

YNIMG Journal 2012 Journal Article

Sparse reduced-rank regression detects genetic associations with voxel-wise longitudinal phenotypes in Alzheimer's disease

• Maria Vounou
• Eva Janousova
• Robin Wolz
• Jason L. Stein
• Paul M. Thompson
• Daniel Rueckert
• Giovanni Montana

Scanning the entire genome in search of variants related to imaging phenotypes holds great promise in elucidating the genetic etiology of neurodegenerative disorders. Here we discuss the application of a penalized multivariate model, sparse reduced-rank regression (sRRR), for the genome-wide detection of markers associated with voxel-wise longitudinal changes in the brain caused by Alzheimer's disease (AD). Using a sample from the Alzheimer's Disease Neuroimaging Initiative database, we performed three separate studies that each compared two groups of individuals to identify genes associated with disease development and progression. For each comparison we took a two-step approach: initially, using penalized linear discriminant analysis, we identified voxels that provide an imaging signature of the disease with high classification accuracy; then we used this multivariate biomarker as a phenotype in a genome-wide association study, carried out using sRRR. The genetic markers were ranked in order of importance of association to the phenotypes using a data re-sampling approach. Our findings confirmed the key role of the APOE and TOMM40 genes but also highlighted some novel potential associations with AD.

YNIMG Journal 2011 Journal Article

A dynamic 4D probabilistic atlas of the developing brain

• Maria Kuklisova-Murgasova
• Paul Aljabar
• Latha Srinivasan
• Serena J. Counsell
• Valentina Doria
• Ahmed Serag
• Ioannis S. Gousias
• James P. Boardman

Probabilistic atlases are widely used in the neuroscience community as a tool for providing a standard space for comparison of subjects and as tissue priors used to enhance the intensity-based classification of brain MRI. Most efforts so far have focused on static brain atlases either for adult or pediatric cohorts. In contrast to the adult brain the rapid growth of the neonatal brain requires an age-specific spatial probabilistic atlas to provide suitable anatomical and structural information. In this paper we describe a 4D probabilistic atlas that allows dynamic generation of prior tissue probability maps for any chosen stage of neonatal brain development between 29 and 44 gestational weeks. The atlas is created from the segmentations of 142 neonatal subjects at different ages using a kernel-based regression method and provides prior tissue probability maps for six structures — cortex, white matter, subcortical grey matter, brainstem, cerebellum and cerebro-spinal fluid. The atlas is publicly available at www. brain-development. org.

YNIMG Journal 2011 Journal Article

Automatic morphometry in Alzheimer's disease and mild cognitive impairment

• Rolf A. Heckemann
• Shiva Keihaninejad
• Paul Aljabar
• Katherine R. Gray
• Casper Nielsen
• Daniel Rueckert
• Joseph V. Hajnal
• Alexander Hammers

This paper presents a novel, publicly available repository of anatomically segmented brain images of healthy subjects as well as patients with mild cognitive impairment and Alzheimer's disease. The underlying magnetic resonance images have been obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. T1-weighted screening and baseline images (1. 5T and 3T) have been processed with the multi-atlas based MAPER procedure, resulting in labels for 83 regions covering the whole brain in 816 subjects. Selected segmentations were subjected to visual assessment. The segmentations are self-consistent, as evidenced by strong agreement between segmentations of paired images acquired at different field strengths (Jaccard coefficient: 0. 802±0. 0146). Morphometric comparisons between diagnostic groups (normal; stable mild cognitive impairment; mild cognitive impairment with progression to Alzheimer's disease; Alzheimer's disease) showed highly significant group differences for individual regions, the majority of which were located in the temporal lobe. Additionally, significant effects were seen in the parietal lobe. Increased left/right asymmetry was found in posterior cortical regions. An automatically derived white-matter hypointensities index was found to be a suitable means of quantifying white-matter disease. This repository of segmentations is a potentially valuable resource to researchers working with ADNI data.

YNIMG Journal 2011 Journal Article

Fast and robust extraction of hippocampus from MR images for diagnostics of Alzheimer's disease

• Jyrki Lötjönen
• Robin Wolz
• Juha Koikkalainen
• Valtteri Julkunen
• Lennart Thurfjell
• Roger Lundqvist
• Gunhild Waldemar
• Hilkka Soininen

Assessment of temporal lobe atrophy from magnetic resonance images is a part of clinical guidelines for the diagnosis of prodromal Alzheimer's disease. As hippocampus is known to be among the first areas affected by the disease, fast and robust definition of hippocampus volume would be of great importance in the clinical decision making. We propose a method for computing automatically the volume of hippocampus using a modified multi-atlas segmentation framework, including an improved initialization of the framework and the correction of partial volume effect. The method produced a high similarity index, 0. 87, and correlation coefficient, 0. 94, with semi-automatically generated segmentations. When comparing hippocampus volumes extracted from 1. 5T and 3T images, the absolute value of the difference was low: 3. 2% of the volume. The correct classification rate for Alzheimer's disease and cognitively normal cases was about 80% while the accuracy 65% was obtained for classifying stable and progressive mild cognitive impairment cases. The method was evaluated in three cohorts consisting altogether about 1000 cases, the main emphasis being in the analysis of the ADNI cohort. The computation time of the method is about 2 minutes on a standard laptop computer. The results show a clear potential for applying the method in clinical practice.

YNIMG Journal 2011 Journal Article

Multi-template tensor-based morphometry: Application to analysis of Alzheimer's disease

• Juha Koikkalainen
• Jyrki Lötjönen
• Lennart Thurfjell
• Daniel Rueckert
• Gunhild Waldemar
• Hilkka Soininen

In this paper methods for using multiple templates in tensor-based morphometry (TBM) are presented and compared to the conventional single-template approach. TBM analysis requires non-rigid registrations which are often subject to registration errors. When using multiple templates and, therefore, multiple registrations, it can be assumed that the registration errors are averaged and eventually compensated. Four different methods are proposed for multi-template TBM. The methods were evaluated using magnetic resonance (MR) images of healthy controls, patients with stable or progressive mild cognitive impairment (MCI), and patients with Alzheimer's disease (AD) from the ADNI database (N =772). The performance of TBM features in classifying images was evaluated both quantitatively and qualitatively. Classification results show that the multi-template methods are statistically significantly better than the single-template method. The overall classification accuracy was 86. 0% for the classification of control and AD subjects, and 72. 1% for the classification of stable and progressive MCI subjects. The statistical group-level difference maps produced using multi-template TBM were smoother, formed larger continuous regions, and had larger t-values than the maps obtained with single-template TBM.

YNIMG Journal 2010 Journal Article

An optimised tract-based spatial statistics protocol for neonates: Applications to prematurity and chronic lung disease

• Gareth Ball
• Serena J. Counsell
• Mustafa Anjari
• Nazakat Merchant
• Tomoki Arichi
• Valentina Doria
• Mary A. Rutherford
• A. David Edwards

Preterm birth is associated with altered white matter microstructure, defined by metrics derived from diffusion tensor imaging (DTI). Tract-based spatial statistics (TBSS) is a useful tool for investigating developing white matter using DTI, but standard TBSS protocols have limitations for neonatal studies. We describe an optimised TBSS protocol for neonatal DTI data, in which registration errors are reduced. As chronic lung disease (CLD) is an independent risk factor for abnormal white matter development, we investigate the effect of this condition on white matter anisotropy and diffusivity using the optimised protocol in a proof of principle experiment. DTI data were acquired from 93 preterm infants (48 male) with a median gestational age at birth of 28+5 (23+4–35+2)weeks at a median postmenstrual age at scan of 41+4 (38+1–46+6)weeks. Nineteen infants developed CLD, defined as requiring supplemental oxygen at 36weeks postmenstrual age. TBSS was modified to include an initial low degrees-of-freedom linear registration step and a second registration to a population-average FA map. The additional registration steps reduced global misalignment between neonatal fractional anisotropy (FA) maps. Infants with CLD had significantly increased radial diffusivity (RD) and significantly reduced FA within the centrum semiovale, corpus callosum and inferior longitudinal fasciculus (p <0. 05) compared to their peers, controlling for degree of prematurity and age at scan. The optimised TBSS protocol improved reliability for neonatal DTI analysis. These data suggest that potentially modifiable respiratory morbidity is associated with widespread altered white matter microstructure in preterm infants at term-equivalent age.

YNIMG Journal 2010 Journal Article

Corrigendum to “An evaluation of four automatic methods of segmenting the subcortical structures in the brain” [NeuroImage 47 (2009) 1435–1447]

• Kolawole Oluwole Babalola
• Brian Patenaude
• Paul Aljabar
• Julia Schnabel
• David Kennedy
• William Crum
• Stephen Smith
• Tim Cootes

YNIMG Journal 2010 Journal Article

Fast and robust multi-atlas segmentation of brain magnetic resonance images

• Jyrki MP. Lötjönen
• Robin Wolz
• Juha R. Koikkalainen
• Lennart Thurfjell
• Gunhild Waldemar
• Hilkka Soininen
• Daniel Rueckert

We introduce an optimised pipeline for multi-atlas brain MRI segmentation. Both accuracy and speed of segmentation are considered. We study different similarity measures used in non-rigid registration. We show that intensity differences for intensity normalised images can be used instead of standard normalised mutual information in registration without compromising the accuracy but leading to threefold decrease in the computation time. We study and validate also different methods for atlas selection. Finally, we propose two new approaches for combining multi-atlas segmentation and intensity modelling based on segmentation using expectation maximisation (EM) and optimisation via graph cuts. The segmentation pipeline is evaluated with two data cohorts: IBSR data (N =18, six subcortial structures: thalamus, caudate, putamen, pallidum, hippocampus, amygdala) and ADNI data (N = 60, hippocampus). The average similarity index between automatically and manually generated volumes was 0. 849 (IBSR, six subcortical structures) and 0. 880 (ADNI, hippocampus). The correlation coefficient for hippocampal volumes was 0. 95 with the ADNI data. The computation time using a standard multicore PC computer was about 3–4 min. Our results compare favourably with other recently published results.

YNIMG Journal 2010 Journal Article

Identifying population differences in whole-brain structural networks: A machine learning approach

• Emma C. Robinson
• Alexander Hammers
• Anders Ericsson
• A. David Edwards
• Daniel Rueckert

Models of whole-brain connectivity are valuable for understanding neurological function, development and disease. This paper presents a machine learning based approach to classify subjects according to their approximated structural connectivity patterns and to identify features which represent the key differences between groups. Brain networks are extracted from diffusion magnetic resonance images obtained by a clinically viable acquisition protocol. Connections are tracked between 83 regions of interest automatically extracted by label propagation from multiple brain atlases followed by classifier fusion. Tracts between these regions are propagated by probabilistic tracking, and mean anisotropy measurements along these connections provide the feature vectors for combined principal component analysis and maximum uncertainty linear discriminant analysis. The approach is tested on two populations with different age distributions: 20–30 and 60–90 years. We show that subjects can be classified successfully (with 87. 46% accuracy) and that the features extracted from the discriminant analysis agree with current consensus on the neurological impact of ageing.

YNIMG Journal 2010 Journal Article

Improving intersubject image registration using tissue-class information benefits robustness and accuracy of multi-atlas based anatomical segmentation

• Rolf A. Heckemann
• Shiva Keihaninejad
• Paul Aljabar
• Daniel Rueckert
• Joseph V. Hajnal
• Alexander Hammers

Automatic anatomical segmentation of magnetic resonance human brain images has been shown to be accurate and robust when based on multiple atlases that encompass the anatomical variability of the cohort of subjects. We observed that the method tends to fail when the segmentation target shows ventricular enlargement that is not captured by the atlas database. By incorporating tissue classification information into the image registration process, we aimed to increase the robustness of the method. For testing, subjects who participated in the Oxford Project to Investigate Memory and Aging (OPTIMA) and the Alzheimer's Disease Neuroimaging Initiative (ADNI) were selected for ventriculomegaly. Segmentation quality was substantially improved in the ventricles and surrounding structures (9/9 successes on visual rating versus 4/9 successes using the baseline method). In addition, the modification resulted in a significant increase of segmentation accuracy in healthy subjects' brain images. Hippocampal segmentation results in a group of patients with temporal lobe epilepsy were near identical with both approaches. The modified approach (MAPER, multi-atlas propagation with enhanced registration) extends the applicability of multi-atlas based automatic whole-brain segmentation to subjects with ventriculomegaly, as seen in normal aging as well as in numerous neurodegenerative diseases.

YNIMG Journal 2010 Journal Article

LEAP: Learning embeddings for atlas propagation

• Robin Wolz
• Paul Aljabar
• Joseph V. Hajnal
• Alexander Hammers
• Daniel Rueckert

We propose a novel framework for the automatic propagation of a set of manually labeled brain atlases to a diverse set of images of a population of subjects. A manifold is learned from a coordinate system embedding that allows the identification of neighborhoods which contain images that are similar based on a chosen criterion. Within the new coordinate system, the initial set of atlases is propagated to all images through a succession of multi-atlas segmentation steps. This breaks the problem of registering images that are very “dissimilar” down into a problem of registering a series of images that are “similar”. At the same time, it allows the potentially large deformation between the images to be modeled as a sequence of several smaller deformations. We applied the proposed method to an exemplar region centered around the hippocampus from a set of 30 atlases based on images from young healthy subjects and a dataset of 796 images from elderly dementia patients and age-matched controls enrolled in the Alzheimer's Disease Neuroimaging Initiative (ADNI). We demonstrate an increasing gain in accuracy of the new method, compared to standard multi-atlas segmentation, with increasing distance between the target image and the initial set of atlases in the coordinate embedding, i. e. , with a greater difference between atlas and image. For the segmentation of the hippocampus on 182 images for which a manual segmentation is available, we achieved an average overlap (Dice coefficient) of 0. 85 with the manual reference.

YNIMG Journal 2010 Journal Article

Measurement of hippocampal atrophy using 4D graph-cut segmentation: Application to ADNI

• Robin Wolz
• Rolf A. Heckemann
• Paul Aljabar
• Joseph V. Hajnal
• Alexander Hammers
• Jyrki Lötjönen
• Daniel Rueckert

We propose a new method of measuring atrophy of brain structures by simultaneously segmenting longitudinal magnetic resonance (MR) images. In this approach a 4D graph is used to represent the longitudinal data: edges are weighted based on spatial and intensity priors and connect spatially and temporally neighboring voxels represented by vertices in the graph. Solving the min-cut/max-flow problem on this graph yields the segmentation for all timepoints in a single step. By segmenting all timepoints simultaneously, a consistent and atrophy-sensitive segmentation is obtained. The application to hippocampal atrophy measurement in 568 image pairs (Baseline and Month 12 follow-up) as well as 362 image triplets (Baseline, Month 12, and Month 24) from the Alzheimer's Disease Neuroimaging Initiative (ADNI) confirms previous findings for atrophy in Alzheimer's disease (AD) and healthy aging. Highly significant correlations between hippocampal atrophy and clinical variables (Mini Mental State Examination, MMSE and Clinical Dementia Rating, CDR) were found and atrophy rates differ significantly according to subjects' ApoE genotype. Based on one year atrophy rates, a correct classification rate of 82% between AD and control subjects is achieved. Subjects that converted from Mild Cognitive Impairment (MCI) to AD after the period for which atrophy was measured (i. e. , after the first 12months) and subjects for whom conversion is yet to be identified were discriminated with a rate of 64%, a promising result with a view to clinical application. Power analysis shows that 67 and 206 subjects are needed for the AD and MCI groups respectively to detect a 25% change in volume loss with 80% power and 5% significance.

YNIMG Journal 2009 Journal Article

An evaluation of four automatic methods of segmenting the subcortical structures in the brain

• Kolawole Oluwole Babalola
• Brian Patenaude
• Paul Aljabar
• Julia Schnabel
• David Kennedy
• William Crum
• Stephen Smith
• Tim Cootes

The automation of segmentation of subcortical structures in the brain is an active research area. We have comprehensively evaluated four novel methods of fully automated segmentation of subcortical structures using volumetric, spatial overlap and distance-based measures. Two methods are atlas-based — classifier fusion and labelling (CFL) and expectation–maximisation segmentation using a brain atlas (EMS), and two incorporate statistical models of shape and appearance — profile active appearance models (PAM) and Bayesian appearance models (BAM). Each method was applied to the segmentation of 18 subcortical structures in 270 subjects from a diverse pool varying in age, disease, sex and image acquisition parameters. Our results showed that all four methods perform on par with recently published methods. CFL performed better than the others according to all three classes of metrics. In summary over all structures, the ranking by the Dice coefficient was CFL, BAM, joint EMS and PAM. The Hausdorff distance ranked the methods as CFL, joint PAM and BAM, EMS, whilst percentage absolute volumetric difference ranked them as joint CFL and PAM, joint BAM and EMS. Furthermore, as we had four methods of performing segmentation, we investigated whether the results obtained by each method were more similar to each other than to the manual segmentations using Williams' Index. Reassuringly, the Williams' Index was close to 1 for most subjects (mean=1. 02, sd=0. 05), indicating better agreement of each method with the gold standard than with the other methods. However, 2% of cases (mainly amygdala and nucleus accumbens) had values outside 3 standard deviations of the mean.

YNIMG Journal 2009 Journal Article

Analysis of serial magnetic resonance images of mouse brains using image registration

• Satheesh Maheswaran
• Hervé Barjat
• Simon T. Bate
• Paul Aljabar
• Derek L.G. Hill
• Lorna Tilling
• Neil Upton
• Michael F. James

The aim of this paper is to investigate techniques that can identify and quantify cross-sectional differences and longitudinal changes in vivo from magnetic resonance images of murine models of brain disease. Two different approaches have been compared. The first approach is a segmentation-based approach: Each subject at each time point is automatically segmented into a number of anatomical structures using atlas-based segmentation. This allows cross-sectional and longitudinal analyses of group differences on a structure-by-structure basis. The second approach is a deformation-based approach: Longitudinal changes are quantified by the registration of each subject's follow-up images to that subject's baseline image. In addition the baseline images can be registered to an atlas allowing voxel-wise analysis of cross-sectional differences between groups. Both approaches have been tested on two groups of mice: A transgenic model of Alzheimer's disease and a wild-type background strain, using serial imaging performed over the age range from 6–14 months. We show that both approaches are able to identify longitudinal and cross-sectional differences. However, atlas-based segmentation suffers from the inability to detect differences across populations and across time in regions which are much smaller than the anatomical regions. In contrast to this, the deformation-based approach can detect statistically significant differences in highly localized areas.

YNIMG Journal 2009 Journal Article

Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration

• Arno Klein
• Jesper Andersson
• Babak A. Ardekani
• John Ashburner
• Brian Avants
• Ming-Chang Chiang
• Gary E. Christensen
• D. Louis Collins

All fields of neuroscience that employ brain imaging need to communicate their results with reference to anatomical regions. In particular, comparative morphometry and group analysis of functional and physiological data require coregistration of brains to establish correspondences across brain structures. It is well established that linear registration of one brain to another is inadequate for aligning brain structures, so numerous algorithms have emerged to nonlinearly register brains to one another. This study is the largest evaluation of nonlinear deformation algorithms applied to brain image registration ever conducted. Fourteen algorithms from laboratories around the world are evaluated using 8 different error measures. More than 45, 000 registrations between 80 manually labeled brains were performed by algorithms including: AIR, ANIMAL, ART, Diffeomorphic Demons, FNIRT, IRTK, JRD-fluid, ROMEO, SICLE, SyN, and four different SPM5 algorithms (“SPM2-type” and regular Normalization, Unified Segmentation, and the DARTEL Toolbox). All of these registrations were preceded by linear registration between the same image pairs using FLIRT. One of the most significant findings of this study is that the relative performances of the registration methods under comparison appear to be little affected by the choice of subject population, labeling protocol, and type of overlap measure. This is important because it suggests that the findings are generalizable to new subject populations that are labeled or evaluated using different labeling protocols. Furthermore, we ranked the 14 methods according to three completely independent analyses (permutation tests, one-way ANOVA tests, and indifference-zone ranking) and derived three almost identical top rankings of the methods. ART, SyN, IRTK, and SPM's DARTEL Toolbox gave the best results according to overlap and distance measures, with ART and SyN delivering the most consistently high accuracy across subjects and label sets. Updates will be published on the http: //www. mindboggle. info/papers/ website.

YNIMG Journal 2008 Journal Article

Automatic segmentation of brain MRIs of 2-year-olds into 83 regions of interest

• Ioannis S. Gousias
• Daniel Rueckert
• Rolf A. Heckemann
• Leigh E. Dyet
• James P. Boardman
• A. David Edwards
• Alexander Hammers

Three-dimensional atlases and databases of the brain at different ages facilitate the description of neuroanatomy and the monitoring of cerebral growth and development. Brain segmentation is challenging in young children due to structural differences compared to adults. We have developed a method, based on established algorithms, for automatic segmentation of young children's brains into 83 regions of interest (ROIs), and applied this to an exemplar group of 33 2-year-old subjects who had been born prematurely. The algorithm uses prior information from 30 normal adult brain magnetic resonance (MR) images, which had been manually segmented to create 30 atlases, each labeling 83 anatomical structures. Each of these adult atlases was registered to each 2-year-old target MR image using non-rigid registration based on free-form deformations. Label propagation from each adult atlas yielded a segmentation of each 2-year-old brain into 83 ROIs. The final segmentation was obtained by combination of the 30 propagated adult atlases using decision fusion, improving accuracy over individual propagations. We validated this algorithm by comparing the automatic approach with three representative manually segmented volumetric regions (the subcortical caudate nucleus, the neocortical pre-central gyrus and the archicortical hippocampus) using similarity indices (SI), a measure of spatial overlap (intersection over average). SI results for automatic versus manual segmentations for these three structures were 0. 90±0. 01, 0. 90±0. 01 and 0. 88±0. 03 respectively. This registration approach allows the rapid construction of automatically labelled age-specific brain atlases for children at the age of 2 years.

YNIMG Journal 2007 Journal Article

Automatic detection and quantification of hippocampal atrophy on MRI in temporal lobe epilepsy: A proof-of-principle study

• Alexander Hammers
• Rolf Heckemann
• Matthias J. Koepp
• John S. Duncan
• Jo V. Hajnal
• Daniel Rueckert
• Paul Aljabar

In temporal lobe epilepsy (TLE), hippocampal atrophy (HA) is a marker of poor prognosis regarding seizure remission, but predicts success of anterior temporal lobe resection. Manual quantification of HA on MRI is time-consuming and limited by investigator availability. Normal ranges of hippocampal volumes, both in absolute terms and relative to intracranial volume, and of hippocampal asymmetry were defined using an automatic label propagation and decision fusion technique based on thirty manually derived atlases of healthy controls. Manual test–retest reliability and overlaps of automatically and manually determined hippocampal volumes were quantified with similarity indices (SIs). Correct clinical identification of ipsilateral HA, and contralaterally normal hippocampal volumes, was determined in nine patients with histologically confirmed hippocampal sclerosis in terms of volumes and asymmetry indices (AIs) for standard statistical thresholds and with receiver operating characteristic (ROC) analysis. Manual test–retest reliability was very high, with SIs between 0. 87 and 0. 90. Manual and automatic hippocampus labels overlapped with a SI of 0. 83 on the unaffected but with 0. 76 on the atrophic side. Accuracy was higher for less atrophic hippocampi. The automatic method correctly identified 6/9 HAs in terms of absolute volume, 7/9 in terms of relative volume at a standard 2 SD threshold, and 9/9 for AIs. ROC-determined thresholds allowed clinically desirable correct identification of all HAs (100% sensitivity) with 85–100% specificity for volumes, and 100% specificity for AIs. The method has the potential to automatically detect unilateral HA, but further work is needed to determine its performance in detecting clinically important bilateral disease.

YNIMG Journal 2007 Journal Article

Automatic segmentation and reconstruction of the cortex from neonatal MRI

• Hui Xue
• Latha Srinivasan
• Shuzhou Jiang
• Mary Rutherford
• A. David Edwards
• Daniel Rueckert
• Joseph V. Hajnal

Segmentation and reconstruction of cortical surfaces from magnetic resonance (MR) images are more challenging for developing neonates than adults. This is mainly due to the dynamic changes in the contrast between gray matter (GM) and white matter (WM) in both T1- and T2-weighted images (T1w and T2w) during brain maturation. In particular in neonatal T2w images WM typically has higher signal intensity than GM. This causes mislabeled voxels during cortical segmentation, especially in the cortical regions of the brain and in particular at the interface between GM and cerebrospinal fluid (CSF). We propose an automatic segmentation algorithm detecting these mislabeled voxels and correcting errors caused by partial volume effects. Our results show that the proposed algorithm corrects errors in the segmentation of both GM and WM compared to the classic expectation maximization (EM) scheme. Quantitative validation against manual segmentation demonstrates good performance (the mean Dice value: 0. 758±0. 037 for GM and 0. 794±0. 078 for WM). The inner, central and outer cortical surfaces are then reconstructed using implicit surface evolution. A landmark study is performed to verify the accuracy of the reconstructed cortex (the mean surface reconstruction error: 0. 73 mm for inner surface and 0. 63 mm for the outer). Both segmentation and reconstruction have been tested on 25 neonates with the gestational ages ranging from ∼27 to 45 weeks. This preliminary analysis confirms previous findings that cortical surface area and curvature increase with age, and that surface area scales to cerebral volume according to a power law, while cortical thickness is not related to age or brain growth.

YNIMG Journal 2006 Journal Article

Abnormal deep grey matter development following preterm birth detected using deformation-based morphometry

• James P. Boardman
• Serena J. Counsell
• Daniel Rueckert
• Olga Kapellou
• Kanwal K. Bhatia
• Paul Aljabar
• Jo Hajnal
• Joanna M. Allsop

Preterm birth is a leading risk factor for neurodevelopmental and cognitive impairment in childhood and adolescence. The most common known cerebral abnormality among preterm infants at term equivalent age is a diffuse white matter abnormality seen on magnetic resonance (MR) images. It occurs with a similar prevalence to subsequent impairment, but its effect on developing neural systems is unknown. MR images were obtained at term equivalent age from 62 infants born at 24–33 completed weeks gestation and 12 term born controls. Tissue damage was quantified using diffusion-weighted imaging, and deformation-based morphometry was used to make a non-subjective survey of the whole brain to identify significant cerebral morphological alterations associated with preterm birth and with diffuse white matter injury. Preterm infants at term equivalent age had reduced thalamic and lentiform volumes without evidence of acute injury in these regions (t = 5. 81, P < 0. 05), and these alterations were more marked with increasing prematurity (t = 7. 13, P < 0. 05 for infants born at less than 28 weeks) and in infants with diffuse white matter injury (t = 6. 43, P < 0. 05). The identification of deep grey matter growth failure in association with diffuse white matter injury suggests that white matter injury is not an isolated phenomenon, but rather, it is associated with the maldevelopment of remote structures. This could be mediated by a disturbance to corticothalamic connectivity during a critical period in cerebral development. Deformation-based morphometry is a powerful tool for modelling the developing brain in health and disease, and can be used to test putative aetiological factors for injury.

YNIMG Journal 2006 Journal Article

Automatic anatomical brain MRI segmentation combining label propagation and decision fusion

• Rolf A. Heckemann
• Joseph V. Hajnal
• Paul Aljabar
• Daniel Rueckert
• Alexander Hammers

Regions in three-dimensional magnetic resonance (MR) brain images can be classified using protocols for manually segmenting and labeling structures. For large cohorts, time and expertise requirements make this approach impractical. To achieve automation, an individual segmentation can be propagated to another individual using an anatomical correspondence estimate relating the atlas image to the target image. The accuracy of the resulting target labeling has been limited but can potentially be improved by combining multiple segmentations using decision fusion. We studied segmentation propagation and decision fusion on 30 normal brain MR images, which had been manually segmented into 67 structures. Correspondence estimates were established by nonrigid registration using free-form deformations. Both direct label propagation and an indirect approach were tested. Individual propagations showed an average similarity index (SI) of 0. 754±0. 016 against manual segmentations. Decision fusion using 29 input segmentations increased SI to 0. 836±0. 009. For indirect propagation of a single source via 27 intermediate images, SI was 0. 779±0. 013. We also studied the effect of the decision fusion procedure using a numerical simulation with synthetic input data. The results helped to formulate a model that predicts the quality improvement of fused brain segmentations based on the number of individual propagated segmentations combined. We demonstrate a practicable procedure that exceeds the accuracy of previous automatic methods and can compete with manual delineations.

YNIMG Journal 2006 Journal Article

Cerebral atrophy measurements using Jacobian integration: Comparison with the boundary shift integral

• Richard G. Boyes
• Daniel Rueckert
• Paul Aljabar
• Jennifer Whitwell
• Jonathan M. Schott
• Derek L.G. Hill
• Nicholas C. Fox

We compared two methods of measuring cerebral atrophy in a cohort of 38 clinically probable Alzheimer's disease (AD) subjects and 22 age-matched normal controls, using metrics of zero atrophy, consistency, scaled atrophy and AD/control group separation. The two methods compared were the boundary shift integral (BSI) and a technique based on the integration of Jacobian determinants from non-rigid registration. For each subject, we used two volumetric magnetic resonance (MR) scans at baseline and a third obtained 1 year later. The case of zero atrophy was established by registering the same-day baseline scan pair, which should approximate zero change. Consistency was established by registering the 1-year follow-up scan to each of the baseline scans, giving two measurements of atrophy that should be very similar, while scaled atrophy was established by reducing one of the same-day scans by a fixed amount, and rigidly registering this to the other same-day scan. Group separation was ascertained by calculating atrophy rates over the two 1-year measures for the control and AD subjects. The results showed the Jacobian integration technique was significantly more accurate in calculating scaled atrophy (P < 0. 001) and was able to distinguish between control and AD subjects more clearly (P < 0. 01).

YNIMG Journal 2006 Journal Article

Tract-based spatial statistics: Voxelwise analysis of multi-subject diffusion data

• Stephen M. Smith
• Mark Jenkinson
• Heidi Johansen-Berg
• Daniel Rueckert
• Thomas E. Nichols
• Clare E. Mackay
• Kate E. Watkins
• Olga Ciccarelli

There has been much recent interest in using magnetic resonance diffusion imaging to provide information about anatomical connectivity in the brain, by measuring the anisotropic diffusion of water in white matter tracts. One of the measures most commonly derived from diffusion data is fractional anisotropy (FA), which quantifies how strongly directional the local tract structure is. Many imaging studies are starting to use FA images in voxelwise statistical analyses, in order to localise brain changes related to development, degeneration and disease. However, optimal analysis is compromised by the use of standard registration algorithms; there has not to date been a satisfactory solution to the question of how to align FA images from multiple subjects in a way that allows for valid conclusions to be drawn from the subsequent voxelwise analysis. Furthermore, the arbitrariness of the choice of spatial smoothing extent has not yet been resolved. In this paper, we present a new method that aims to solve these issues via (a) carefully tuned non-linear registration, followed by (b) projection onto an alignment-invariant tract representation (the “mean FA skeleton”). We refer to this new approach as Tract-Based Spatial Statistics (TBSS). TBSS aims to improve the sensitivity, objectivity and interpretability of analysis of multi-subject diffusion imaging studies. We describe TBSS in detail and present example TBSS results from several diffusion imaging studies.