Author name cluster

Yi Su

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

27 papers

2 author rows

EAAI Journal 2026 Journal Article

Natural gas demand forecasting via latent pattern retrieval and expert specialization

Sichong Lu
Hairun Wang
Jiahui Chai
Yi Su
Lean Yu
Bo Yang

Natural gas demand forecasting is challenged by pattern heterogeneity and recurring cycles, which existing global modeling approaches cannot effectively address. This study proposes a latent pattern retrieval and expert specialization (LPRES) framework that strategically leverages pretrained language models (PLMs) not as direct forecasting tools but through role-specific adaptation. First, latent pattern awareness is developed through a temporal feature learning model derived from a PLM fine-tuned for anomaly detection, exploiting its sensitivity to pattern changes. Second, based on this model, an adaptive sliding window segmentation algorithm partitions historical data into segments, each corresponding to a distinct latent pattern. Third, for each identified latent pattern, specialized forecasting experts are trained using a PLM fine-tuned on large-scale forecasting tasks, thereby adapting its strong predictive capacity to the characteristics of individual latent patterns. Fourth, during forecasting, input windows are matched to their most similar latent pattern through similarity-based retrieval and routed to the corresponding expert. Experiments on four natural gas datasets spanning monthly and hourly frequencies show that LPRES delivers competitive forecasting performance across diverse data characteristics, achieving mean absolute percentage error (MAPE) values from 0. 035 to 0. 107 and reducing errors by up to 11. 2% relative to the strongest baselines. A complementary theoretical framework identifies the conditions under which expert specialization is most beneficial.

EAAI Journal 2025 Journal Article

A novel framework for crack segmentation using image augmentation and a CannyNet

Gang Liu
Xuming Li
Jin Di
Rui Sun
Fengjiang Qin
Yi Su
Dewei Liu

Computer-vision based crack detection is highly dependent on the quality of the segmentation process and remains a challenging task due to its complexity. In this paper, a framework for image segmentation that incorporates image augmentation method and a CannyNet is proposed to improve segmentation results. The style transfer is employed for image augmentation. A novel deep neural network for crack image segmentation, named CannyNet, is proposed to enhance the recognition capability for tiny cracks. Moreover, to improve the precision of CannyNet predictions, Bayesian optimization approach is employed to optimize network hyperparameters. The proposed framework for crack segmentation was verified using four open-source dataset and a new constructed dataset by conducting experimental test. A comparison of segmentation models indicates that style transfer method enhances the model's generalization, and the CannyNet demonstrates superior performance. The Bayesian optimization strategy is capable of optimizing the architecture of the CannyNet, thereby improving crack segmentation results.

ECAI Conference 2025 Conference Paper

DetTrack: Realizing Strong Identity Preservation in Multi-Object Tracking via exploration of Detection Information

Yi Zhang
Yi Su
Chen Luo

Multiple Object Tracking (MOT) aims to detect all objects in the scene and associate them across frames with unique ID. Within tracking-by-detection (TBD) paradigm, the confidence based two-stage matching scheme has become popular in MOT. However, when two detections are matched to the same trajectory, the one with higher confidence score usually takes precedence over the lower one, even if the lower one is the ground-truth, causing ID switches (IDS). Considering this, we propose a tailored filtering mechanism to handle the low-confident detections in a more reasonable way. Besides, we introduce a novel fusion scheme for appearance and motion information based on appearance clarity and localization accuracy of the detection boxes. Finally, an adaptive management of unmatched detections scheme is proposed to reduce the occurrence of IDS and duplicate trajectories. Extensive experiments have been conducted on MOT17 and MOT20, in which our tracker exhibits stronger identity preservation capabilities against other competitors.

ICLR Conference 2025 Conference Paper

Training Language Models to Self-Correct via Reinforcement Learning

Aviral Kumar
Vincent Zhuang
Rishabh Agarwal
Yi Su
John D. Co-Reyes
Avi Singh
Kate Baumli
Shariq Iqbal

Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.

NeurIPS Conference 2024 Conference Paper

Online Feature Updates Improve Online (Generalized) Label Shift Adaptation

Ruihan Wu
Siddhartha Datta
Yi Su
Dheeraj Baby
Yu-Xiang Wang
Kilian Q. Weinberger

This paper addresses the prevalent issue of label shift in an online setting with missing labels, where data distributions change over time and obtaining timely labels is challenging. While existing methods primarily focus on adjusting or updating the final layer of a pre-trained classifier, we explore the untapped potential of enhancing feature representations using unlabeled data at test-time. Our novel method, Online Label Shift adaptation with Online Feature Updates (OLS-OFU), leverages self-supervised learning to refine the feature extraction process, thereby improving the prediction model. By carefully designing the algorithm, theoretically OLS-OFU maintains the similar online regret convergence to the results in the literature while taking the improved features into account. Empirically, it achieves substantial improvements over existing methods, which is as significant as the gains existing methods have over the baseline (i. e. , without distribution shift adaptations).

PDF Details DOI

YNICL Journal 2024 Journal Article

Predicting cognitive decline: Which is more useful, baseline amyloid levels or longitudinal change?

Gengsheng Chen
Nicole S. McKay
Brian A. Gordon
Jingxia Liu
Nelly Joseph-Mathurin
Suzanne E. Schindler
Jason Hassenstab
Andrew J. Aschenbrenner

C-Pittsburgh compound B (PiB) Aβ-PET to predict cognitive decline. A cohort of 153 participants who previously underwent PiB-PET scans and comprehensive clinical assessments were used in this study. Our analyses revealed that baseline Aβ is significantly associated with the rate of change in cognitive composite scores, with cognition declining more rapidly when baseline PiB Aβ levels were higher. In contrast, no signification association was identified between the rate of change in PiB-PET Aβ and cognitive decline. Additionally, the ability of the rate of change in the PiB-PET measures to predict cognitive decline was significantly influenced by APOE ε4 carrier status. These results suggest that a single PiB-PET scan is sufficient to predict cognitive decline and that longitudinal measures of Aβ accumulation do not improve the prediction of cognitive decline once someone is amyloid positive.

YNICL Journal 2023 Journal Article

Molecular imaging of the association between serotonin degeneration and beta-amyloid deposition in mild cognitive impairment

Gwenn S. Smith
Hillary Protas
Hiroto Kuwabara
Alena Savonenko
Najlla Nassery
Neda F. Gould
Michael Kraut
Dimitri Avramopoulos

BACKGROUND: Degeneration of the serotonin system has been observed in Alzheimer's disease (AD) and in mild cognitive impairment (MCI). In transgenic amyloid mouse models, serotonin degeneration is detected prior to widespread cortical beta-amyloid (Aβ) deposition, also suggesting that serotonin degeneration may be observed in preclinical AD. METHODS: The differences in the distribution of serotonin degeneration (reflected by the loss of the serotonin transporter, 5-HTT) relative to Aβ deposition was measured with positron emission tomography in a group of individuals with MCI and a group of healthy older adults. A multi-modal partial least squares (mmPLS) algorithm was applied to identify the spatial covariance pattern between 5-HTT availability and Aβ deposition. RESULTS: Forty-five individuals with MCI and 35 healthy older adults were studied, 22 and 27 of whom were included in the analyses who were "amyloid positive" and "amyloid negative", respectively. A pattern of lower cortical, subcortical and limbic 5-HTT availability and higher cortical Aβ deposition distinguished the MCI from the healthy older control participants. Greater expression of this pattern was correlated with greater deficits in memory and executive function in the MCI group, not in the control group. CONCLUSION: A spatial covariance pattern of lower 5-HTT availability and Aβ deposition was observed to a greater extent in an MCI group relative to a control group and was associated with cognitive impairment in the MCI group. The results support the application of mmPLS to understand the neurochemical changes associated with Aβ deposition in the course of preclinical AD.

ICLR Conference 2023 Conference Paper

Offline RL for Natural Language Generation with Implicit Language Q Learning

Charlie Victor Snell
Ilya Kostrikov
Yi Su
Sherry Yang 0001
Sergey Levine

Large language models distill broad knowledge from text corpora. However, they can be inconsistent when it comes to completing user specified tasks. This issue can be addressed by finetuning such models via supervised learning on curated datasets, or via reinforcement learning. In this work, we propose a novel offline RL method, implicit language Q-learning (ILQL), designed for use on language models, that combines both the flexible utility maximization framework of RL algorithms with the ability of supervised learning to leverage previously collected data, as well as its simplicity and stability. Our method employs a combination of value conservatism alongside an implicit dataset support constraint in learning value functions, which are then used to guide language model generations towards maximizing user-specified utility functions. In addition to empirically validating ILQL, we present a detailed empirical analysis of situations where offline RL can be useful in natural language generation settings, demonstrating how it can be a more effective utility optimizer than prior approaches for end-to-end dialogue, and how it can effectively optimize high variance reward functions based on subjective judgement, such as whether to label a comment as toxic or not.

NeurIPS Conference 2023 Conference Paper

Unified Off-Policy Learning to Rank: a Reinforcement Learning Perspective

Zeyu Zhang
Yi Su
Hui Yuan
Yiran Wu
Rishab Balasubramanian
Qingyun Wu
Huazheng Wang
Mengdi Wang

Off-policy Learning to Rank (LTR) aims to optimize a ranker from data collected by a deployed logging policy. However, existing off-policy learning to rank methods often make strong assumptions about how users generate the click data, i. e. , the click model, and hence need to tailor their methods specifically under different click models. In this paper, we unified the ranking process under general stochastic click models as a Markov Decision Process (MDP), and the optimal ranking could be learned with offline reinforcement learning (RL) directly. Building upon this, we leverage offline RL techniques for off-policy LTR and propose the Click Model-Agnostic Unified Off-policy Learning to Rank (CUOLR) method, which could be easily applied to a wide range of click models. Through a dedicated formulation of the MDP, we show that offline RL algorithms can adapt to various click models without complex debiasing techniques and prior knowledge of the model. Results on various large-scale datasets demonstrate that CUOLR consistently outperforms the state-of-the-art off-policy learning to rank algorithms while maintaining consistency and robustness under different click models.

NeurIPS Conference 2022 Conference Paper

Data-Driven Offline Decision-Making via Invariant Representation Learning

Han Qi
Yi Su
Aviral Kumar
Sergey Levine

The goal in offline data-driven decision-making is synthesize decisions that optimize a black-box utility function, using a previously-collected static dataset, with no active interaction. These problems appear in many forms: offline reinforcement learning (RL), where we must produce actions that optimize the long-term reward, bandits from logged data, where the goal is to determine the correct arm, and offline model-based optimization (MBO) problems, where we must find the optimal design provided access to only a static dataset. A key challenge in all these settings is distributional shift: when we optimize with respect to the input into a model trained from offline data, it is easy to produce an out-of-distribution (OOD) input that appears erroneously good. In contrast to prior approaches that utilize pessimism or conservatism to tackle this problem, in this paper, we formulate offline data-driven decision-making as domain adaptation, where the goal is to make accurate predictions for the value of optimized decisions (“target domain”), when training only on the dataset (“source domain”). This perspective leads to invariant objective models (IOM), our approach for addressing distributional shift by enforcing invariance between the learned representations of the training dataset and optimized decisions. In IOM, if the optimized decisions are too different from the training dataset, the representation will be forced to lose much of the information that distinguishes good designs from bad ones, making all choices seem mediocre. Critically, when the optimizer is aware of this representational tradeoff, it should choose not to stray too far from the training distribution, leading to a natural trade-off between distributional shift and learning performance.

JMLR Journal 2022 Journal Article

Tianshou: A Highly Modularized Deep Reinforcement Learning Library

Jiayi Weng
Huayu Chen
Dong Yan
Kaichao You
Alexis Duburcq
Minghao Zhang
Yi Su
Hang Su

In this paper, we present Tianshou, a highly modularized Python library for deep reinforcement learning (DRL) that uses PyTorch as its backend. Tianshou intends to be research-friendly by providing a flexible and reliable infrastructure of DRL algorithms. It supports online and offline training with more than 20 classic algorithms through a unified interface. To facilitate related research and prove Tianshou's reliability, we have released Tianshou's benchmark of MuJoCo environments, covering eight classic algorithms with state-of-the-art performance. We open-sourced Tianshou at https://github.com/thu-ml/tianshou/. [abs] [ pdf ][ bib ] [ code ] &copy JMLR 2022. ( edit, beta )

NeurIPS Conference 2021 Conference Paper

Online Adaptation to Label Distribution Shift

Ruihan Wu
Chuan Guo
Yi Su
Kilian Q. Weinberger

Machine learning models often encounter distribution shifts when deployed in the real world. In this paper, we focus on adaptation to label distribution shift in the online setting, where the test-time label distribution is continually changing and the model must dynamically adapt to it without observing the true label. This setting is common in many real world scenarios such as medical diagnosis, where disease prevalences can vary substantially at different times of the year. Leveraging a novel analysis, we show that the lack of true label does not hinder estimation of the expected test loss, which enables the reduction of online label shift adaptation to conventional online learning. Informed by this observation, we propose adaptation algorithms inspired by classical online learning techniques such as Follow The Leader (FTL) and Online Gradient Descent (OGD) and derive their regret bounds. We empirically verify our findings under both simulated and real world label distribution shifts and show that OGD is particularly effective and robust to a variety of challenging label shift scenarios.

YNICL Journal 2021 Journal Article

PET evidence of preclinical cerebellar amyloid plaque deposition in autosomal dominant Alzheimer’s disease-causing Presenilin-1 E280A mutation carriers

Valentina Ghisays
Francisco Lopera
Dhruman D. Goradia
Hillary D. Protas
Michael H. Malek-Ahmadi
Yinghua Chen
Vivek Devadas
Ji Luo

BACKGROUND: In contrast to sporadic Alzheimer's disease, autosomal dominant Alzheimer's disease (ADAD) is associated with greater neuropathological evidence of cerebellar amyloid plaque (Aβ) deposition. In this study, we used positron emission tomography (PET) measurements of fibrillar Aβ burden to characterize the presence and age at onset of cerebellar Aβ deposition in cognitively unimpaired (CU) Presenilin-1 (PSEN1) E280A mutation carriers from the world's largest extended family with ADAD. METHODS: C Pittsburgh compound B (PiB) PET data from two independent studies - API ADAD Colombia Trial (NCT01998841) and Colombia-Boston (COLBOS) longitudinal biomarker study were included. The tracers were selected independently by the respective sponsors prior to the start of each study and used exclusively throughout. Template-based cerebellar Aβ-SUVR (standard-uptake value ratios) using a known-to-be-spared pons reference region (cerebellar SUVR_pons), to a) compare 28-56-year-old CU carriers and non-carriers; b) estimate the age at which cerebellar SUVR_pons began to differ significantly in carrier and non-carrier groups; and c) characterize in carriers associations with age, cortical SUVR_pons, delayed recall memory, and API ADAD composite score. RESULTS: Florbetapir and PiB cerebellar SUVR_pons were significantly higher in carriers than non-carriers (p < 0.0001). Cerebellar SUVR_pons began to distinguish carriers from non-carriers at age 34, 10 years before the carriers' estimated age at mild cognitive impairment onset. Florbetapir and PiB cerebellar SUVR_pons in carriers were positively correlated with age (r = 0.44 & 0.69, p < 0.001), cortical SUVR_pons (r = 0.55 & 0.69, p < 0.001), and negatively correlated with delayed recall memory (r = -0.21 & -0.50, p < 0.05, unadjusted for cortical SUVR_pons) and API ADAD composite (r = -0.25, p < 0.01, unadjusted for cortical SUVR_pons in florbetapir API ADAD cohort). CONCLUSION: This PET study provides evidence of cerebellar Aβ plaque deposition in CU carriers starting about a decade before the clinical onset of ADAD. Additional studies are needed to clarify the impact of using a cerebellar versus pons reference region on the power to detect and track ADAD changes, even in preclinical stages of this disorder.

YNICL Journal 2020 Journal Article

AD-NET: Age-adjust neural network for improved MCI to AD conversion prediction

Fei Gao
Hyunsoo Yoon
Yanzhe Xu
Dhruman Goradia
Ji Luo
Teresa Wu
Yi Su

The prediction of Mild Cognitive Impairment (MCI) patients who are at higher risk converting to Alzheimer's Disease (AD) is critical for effective intervention and patient selection in clinical trials. Different biomarkers including neuroimaging have been developed to serve the purpose. With extensive methodology development efforts on neuroimaging, an emerging field is deep learning research. One great challenge facing deep learning is the limited medical imaging data available. To address the issue, researchers explore the use of transfer learning to extend the applicability of deep models on neuroimaging research for AD diagnosis and prognosis. Existing transfer learning models mostly focus on transferring the features from the pre-training into the fine-tuning stage. Recognizing the advantages of the knowledge gained during the pre-training, we propose an AD-NET (Age-adjust neural network) with the pre-training model serving two purposes: extracting and transferring features; and obtaining and transferring knowledge. Specifically, the knowledge being transferred in this research is an age-related surrogate biomarker. To evaluate the effectiveness of the proposed approach, AD-NET is compared with 8 classification models from literature using the same public neuroimaging dataset. Experimental results show that the proposed AD-NET outperforms the competing models in predicting the MCI patients at risk for conversion to the AD stage.

ICML Conference 2020 Conference Paper

Adaptive Estimator Selection for Off-Policy Evaluation

Yi Su
Pavithra Srinath
Akshay Krishnamurthy

We develop a generic data-driven method for estimator selection in off-policy policy evaluation settings. We establish a strong performance guarantee for the method, showing that it is competitive with the oracle estimator, up to a constant factor. Via in-depth case studies in contextual bandits and reinforcement learning, we demonstrate the generality and applicability of the method. We also perform comprehensive experiments, demonstrating the empirical efficacy of our approach and comparing with related approaches. In both case studies, our method compares favorably with existing methods.

YNICL Journal 2020 Journal Article

Applying surface-based morphometry to study ventricular abnormalities of cognitively unimpaired subjects prior to clinically significant memory decline

Qunxi Dong
Wen Zhang
Cynthia M. Stonnington
Jianfeng Wu
Boris A. Gutman
Kewei Chen
Yi Su
Leslie C. Baxter

Ventricular volume (VV) is a widely used structural magnetic resonance imaging (MRI) biomarker in Alzheimer's disease (AD) research. Abnormal enlargements of VV can be detected before clinically significant memory decline. However, VV does not pinpoint the details of subregional ventricular expansions. Here we introduce a ventricular morphometry analysis system (VMAS) that generates a whole connected 3D ventricular shape model and encodes a great deal of ventricular surface deformation information that is inaccessible by VV. VMAS contains an automated segmentation approach and surface-based multivariate morphometry statistics. We applied VMAS to two independent datasets of cognitively unimpaired (CU) groups. To our knowledge, it is the first work to detect ventricular abnormalities that distinguish normal aging subjects from those who imminently progress to clinically significant memory decline. Significant bilateral ventricular morphometric differences were first shown in 38 members of the Arizona APOE cohort, which included 18 CU participants subsequently progressing to the clinically significant memory decline within 2 years after baseline visits (progressors), and 20 matched CU participants with at least 4 years of post-baseline cognitive stability (non-progressors). VMAS also detected significant differences in bilateral ventricular morphometry in 44 Alzheimer's Disease Neuroimaging Initiative (ADNI) subjects (18 CU progressors vs. 26 CU non-progressors) with the same inclusion criterion. Experimental results demonstrated that the ventricular anterior horn regions were affected bilaterally in CU progressors, and more so on the left. VMAS may track disease progression at subregional levels and measure the effects of pharmacological intervention at a preclinical stage.

YNICL Journal 2020 Journal Article

Comparing cortical signatures of atrophy between late-onset and autosomal dominant Alzheimer disease

Aylin Dincer
Brian A. Gordon
Amrita Hari-Raj
Sarah J. Keefe
Shaney Flores
Nicole S. McKay
Angela M. Paulick
Kristine E. Shady Lewis

F-florbetapir. To generate cortical signature maps of cortical thickness, we performed a vertex-wise analysis between the cognitively normal controls and impaired groups within each cohort using six increasingly conservative statistical thresholds to determine significance. The optimal cortical map among the six statistical thresholds was determined from a receiver operating characteristic analysis testing the performance of each map in discriminating between the cognitively normal controls and preclinical groups. We then performed within-cohort and cross-cohort (e.g. ADAD maps evaluated in the Knight ADRC cohort) analyses to examine the sensitivity of the optimal cortical signature maps to the amyloid levels using only the cognitively normal individuals (cognitively normal controls and preclinical groups) in comparison to hippocampal volume. We found the optimal cortical signature maps were sensitive to early increases in amyloid for the asymptomatic individuals within their respective cohorts and were significant beyond the inclusion of hippocampus volume, but the cortical signature maps performed poorly when analyzing across cohorts. These results suggest the cortical signature maps are a useful MRI biomarker of early AD-related neurodegeneration in preclinical individuals and the pattern of decline differs between LOAD and ADAD.

ICML Conference 2020 Conference Paper

Doubly robust off-policy evaluation with shrinkage

Yi Su
Maria Dimakopoulou
Akshay Krishnamurthy
Miroslav Dudík

We propose a new framework for designing estimators for off-policy evaluation in contextual bandits. Our approach is based on the asymptotically optimal doubly robust estimator, but we shrink the importance weights to minimize a bound on the mean squared error, which results in a better bias-variance tradeoff in finite samples. We use this optimization-based framework to obtain three estimators: (a) a weight-clipping estimator, (b) a new weight-shrinkage estimator, and (c) the first shrinkage-based estimator for combinatorial action sets. Extensive experiments in both standard and combinatorial bandit benchmark problems show that our estimators are highly adaptive and typically outperform state-of-the-art methods.

ICML Conference 2019 Conference Paper

CAB: Continuous Adaptive Blending for Policy Evaluation and Learning

Yi Su
Lequn Wang
Michele Santacatterina
Thorsten Joachims

The ability to perform offline A/B-testing and off-policy learning using logged contextual bandit feedback is highly desirable in a broad range of applications, including recommender systems, search engines, ad placement, and personalized health care. Both offline A/B-testing and off-policy learning require a counterfactual estimator that evaluates how some new policy would have performed, if it had been used instead of the logging policy. In this paper, we identify a family of counterfactual estimators which subsumes most such estimators proposed to date. Our analysis of this family identifies a new estimator - called Continuous Adaptive Blending (CAB) - which enjoys many advantageous theoretical and practical properties. In particular, it can be substantially less biased than clipped Inverse Propensity Score (IPS) weighting and the Direct Method, and it can have less variance than Doubly Robust and IPS estimators. In addition, it is sub-differentiable such that it can be used for learning, unlike the SWITCH estimator. Experimental results show that CAB provides excellent evaluation accuracy and outperforms other counterfactual estimators in terms of learning performance.

YNICL Journal 2019 Journal Article

Quantification of white matter cellularity and damage in preclinical and early symptomatic Alzheimer's disease

Qing Wang
Yong Wang
Jingxia Liu
Courtney L. Sutphen
Carlos Cruchaga
Tyler Blazey
Brian A. Gordon
Yi Su

Interest in understanding the roles of white matter (WM) inflammation and damage in the pathophysiology of Alzheimer disease (AD) has been growing significantly in recent years. However, in vivo magnetic resonance imaging (MRI) techniques for imaging inflammation are still lacking. An advanced diffusion-based MRI method, neuro-inflammation imaging (NII), has been developed to clinically image and quantify WM inflammation and damage in AD. Here, we employed NII measures in conjunction with cerebrospinal fluid (CSF) biomarker classification (for β-amyloid (Aβ) and neurodegeneration) to evaluate 200 participants in an ongoing study of memory and aging. Elevated NII-derived cellular diffusivity was observed in both preclinical and early symptomatic phases of AD, while disruption of WM integrity, as detected by decreased fractional anisotropy (FA) and increased radial diffusivity (RD), was only observed in the symptomatic phase of AD. This may suggest that WM inflammation occurs earlier than WM damage following abnormal Aβ accumulation in AD. The negative correlation between NII-derived cellular diffusivity and CSF Aβ42 level (a marker of amyloidosis) may indicate that WM inflammation is associated with increasing Aβ burden. NII-derived FA also negatively correlated with CSF t-tau level (a marker of neurodegeneration), suggesting that disruption of WM integrity is associated with increasing neurodegeneration. Our findings demonstrated the capability of NII to simultaneously image and quantify WM cellularity changes and damage in preclinical and early symptomatic AD. NII may serve as a clinically feasible imaging tool to study the individual and composite roles of WM inflammation and damage in AD.

YNICL Journal 2018 Journal Article

Utilizing the Centiloid scale in cross-sectional and longitudinal PiB PET studies

Yi Su
Shaney Flores
Russ C. Hornbeck
Benjamin Speidel
Andrei G. Vlassenko
Brian A. Gordon
Robert A. Koeppe
William E. Klunk

Amyloid imaging is a valuable tool for research and diagnosis in dementing disorders. Successful use of this tool is limited by the lack of a common standard in the quantification of amyloid imaging data. The Centiloid approach was recently proposed to address this problem and in this work, we report our implementation of this approach and evaluate the impact of differences in underlying image analysis methodologies using both cross-sectional and longitudinal datasets. The Centiloid approach successfully converts quantitative amyloid burden measurements into a common Centiloid scale (CL) and comparable dynamic range. As expected, the Centiloid values derived from different analytical approaches inherit some of the inherent benefits and drawbacks of the underlying approaches, and these differences result in statistically significant (p < 0. 05) differences in the variability and group mean values. Because of these differences, even after expression in CL, the 95% specificity amyloid positivity thresholds derived from different analytic approaches varied from 5. 7 CL to 11. 9 CL, and the reliable worsening threshold varied from −2. 0 CL to 11. 0 CL. Although this difference is in part due to the dependency of the threshold determination methodology on the statistical characteristics of the measurements. When amyloid measurements obtained from different centers are combined for analysis, one should not expect Centiloid conversion to eliminate all the differences in amyloid burden measurements due to variabilities in underlying acquisition protocols and analysis techniques.

YNIMG Journal 2017 Journal Article

AV-1451 PET imaging of tau pathology in preclinical Alzheimer disease: Defining a summary measure

Shruti Mishra
Brian A. Gordon
Yi Su
Jon Christensen
Karl Friedrichsen
Kelley Jackson
Russ Hornbeck
David A. Balota

Utilizing [18F]-AV-1451 tau positron emission tomography (PET) as an Alzheimer disease (AD) biomarker will require identification of brain regions that are most important in detecting elevated tau pathology in preclinical AD. Here, we utilized an unsupervised learning, data-driven approach to identify brain regions whose tau PET is most informative in discriminating low and high levels of [18F]-AV-1451 binding. 84 cognitively normal participants who had undergone AV-1451 PET imaging were used in a sparse k-means clustering with resampling analysis to identify the regions most informative in dividing a cognitively normal population into high tau and low tau groups. The highest-weighted FreeSurfer regions of interest (ROIs) separating these groups were the entorhinal cortex, amygdala, lateral occipital cortex, and inferior temporal cortex, and an average SUVR in these four ROIs was used as a summary metric for AV-1451 uptake. We propose an AV-1451 SUVR cut-off of 1. 25 to define high tau as described by imaging. This spatial distribution of tau PET is a more widespread pattern than that predicted by pathological staging schemes. Our data-derived metric was validated first in this cognitively normal cohort by correlating with early measures of cognitive dysfunction, and with disease progression as measured by β-amyloid PET imaging. We additionally validated this summary metric in a cohort of 13 Alzheimer disease patients, and showed that this measure correlates with cognitive dysfunction and β-amyloid PET imaging in a diseased population.

JBHI Journal 2016 Journal Article

Volume Preserved Mass–Spring Model with Novel Constraints for Soft Tissue Deformation

Yuping Duan
Weimin Huang
Huibin Chang
Wenyu Chen
Jiayin Zhou
Soo Kng Teo
Yi Su
Chee Kong Chui

An interactive surgical simulation system needs to meet three main requirements, speed, accuracy, and stability. In this paper, we present a stable and accurate method for animating mass–spring systems in real time. An integration scheme derived from explicit integration is used to obtain interactive realistic animation for a multiobject environment. We explore a predictor–corrector approach by correcting the estimation of the explicit integration in a poststep process. We introduce novel constraints on positions into the mass–spring model (MSM) to model the nonlinearity and preserve volume for the realistic simulation of the incompressibility. We verify the proposed MSM by comparing its deformations with the reference deformations of the nonlinear finite-element method. Moreover, experiments on porcine organs are designed for the evaluation of the multiobject deformation. Using a pair of freshly harvested porcine liver and gallbladder, the real organ deformations are acquired by computed tomography and used as the reference ground truth. Compared to the porcine model, our model achieves a $1. 502$ mm mean absolute error measured at landmark locations for cases with small deformation (the largest deformation is $49. 109$ mm) and a $3. 639$ mm mean absolute error for cases with large deformation (the largest deformation is $83. 137$ mm). The changes of volume for the two deformations are limited to $0. 030\%$ and $0. 057\%$, respectively. Finally, an implementation in a virtual reality environment for laparoscopic cholecystectomy demonstrates that our model is capable to simulate large deformation and preserve volume in real-time calculations.

YNIMG Journal 2015 Journal Article

MR-based attenuation correction for PET/MRI neurological studies with continuous-valued attenuation coefficients for bone through a conversion from R2* to CT-Hounsfield units

Meher R. Juttukonda
Bryant G. Mersereau
Yasheng Chen
Yi Su
Brian G. Rubin
Tammie L.S. Benzinger
David S. Lalush
Hongyu An

Aim MR-based correction for photon attenuation in PET/MRI remains challenging, particularly for neurological applications requiring quantitation of data. Existing methods are either not sufficiently accurate or are limited by the computation time required. The goal of this study was to develop an MR-based attenuation correction method that accurately separates bone tissue from air and provides continuous-valued attenuation coefficients for bone. Materials and methods PET/MRI and CT datasets were obtained from 98 subjects (mean age [±SD]: 66yrs [±9. 8], 57 females) using an IRB-approved protocol and with informed consent. Subjects were injected with 352±29MBq of 18F-Florbetapir tracer, and PET acquisitions were begun either immediately or 50min after injection. CT images of the head were acquired separately using a PET/CT system. Dual echo ultrashort echo-time (UTE) images and two-point Dixon images were acquired. Regions of air were segmented via a threshold of the voxel-wise multiplicative inverse of the UTE echo 1 image. Regions of bone were segmented via a threshold of the R2* image computed from the UTE echo 1 and UTE echo 2 images. Regions of fat and soft tissue were segmented using fat and water images decomposed from the Dixon images. Air, fat, and soft tissue were assigned linear attenuation coefficients (LACs) of 0, 0. 092, and 0. 1cm−1, respectively. LACs for bone were derived from a regression analysis between corresponding R2* and CT values. PET images were reconstructed using the gold standard CT method and the proposed CAR-RiDR method. Results The RiDR segmentation method produces mean Dice coefficient±SD across subjects of 0. 75±0. 05 for bone and 0. 60±0. 08 for air. The CAR model for bone LACs greatly improves accuracy in estimating CT values (28. 2%±3. 0 mean error) compared to the use of a constant CT value (46. 9%±5. 8, p<10−6). Finally, the CAR-RiDR method provides a low whole-brain mean absolute percent-error (MAPE±SD) in PET reconstructions across subjects of 2. 55%±0. 86. Regional PET errors were also low and ranged from 0. 88% to 3. 79% in 24 brain ROIs. Conclusion We propose an MR-based attenuation correction method (CAR-RiDR) for quantitative PET neurological imaging. The proposed method employs UTE and Dixon images and consists of two novel components: 1) accurate segmentation of air and bone using the inverse of the UTE1 image and the R2* image, respectively and 2) estimation of continuous LAC values for bone using a regression between R2* and CT-Hounsfield units. From our analysis, we conclude that the proposed method closely approaches (<3% error) the gold standard CT-scaled method in PET reconstruction accuracy.

YNIMG Journal 2015 Journal Article

Partial volume correction in quantitative amyloid imaging

Yi Su
Tyler M. Blazey
Abraham Z. Snyder
Marcus E. Raichle
Daniel S. Marcus
Beau M. Ances
Randall J. Bateman
Nigel J. Cairns

Amyloid imaging is a valuable tool for research and diagnosis in dementing disorders. As positron emission tomography (PET) scanners have limited spatial resolution, measured signals are distorted by partial volume effects. Various techniques have been proposed for correcting partial volume effects, but there is no consensus as to whether these techniques are necessary in amyloid imaging, and, if so, how they should be implemented. We evaluated a two-component partial volume correction technique and a regional spread function technique using both simulated and human Pittsburgh compound B (PiB) PET imaging data. Both correction techniques compensated for partial volume effects and yielded improved detection of subtle changes in PiB retention. However, the regional spread function technique was more accurate in application to simulated data. Because PiB retention estimates depend on the correction technique, standardization is necessary to compare results across groups. Partial volume correction has sometimes been avoided because it increases the sensitivity to inaccuracy in image registration and segmentation. However, our results indicate that appropriate PVC may enhance our ability to detect changes in amyloid deposition.

YNIMG Journal 2015 Journal Article

Preclinical evaluation of a promising C-11 labeled PET tracer for imaging phosphodiesterase 10A in the brain of living subject

Hui Liu
Hongjun Jin
Xuyi Yue
Xiang Zhang
Hao Yang
Junfeng Li
Hubert Flores
Yi Su

Phosphodiesterase 10A (PDE10A) plays a key role in the regulation of brain striatal signaling. A PET tracer for PDE10A may serve as a tool to evaluate PDE10A expression in vivo in central nervous system disorders with striatal pathology. Here, we further characterized the binding properties of a previously reported radioligand we developed for PDE10A, [11C]TZ1964B, in rodents and nonhuman primates (NHPs). The tritiated counterpart [3H]TZ1964B was used for in vitro binding characterizations in rat striatum homogenates and in vitro autoradiographic studies in rat brain slices. The carbon-11 labeled [11C]TZ1964B was utilized in the ex vivo autoradiography studies for the brain of rats and microPET imaging studies for the brain of NHPs. MicroPET scans of [11C]TZ1964B in NHPs were conducted at baseline, as well as with using a selective PDE10A inhibitor MP-10 for either pretreatment or displacement. The in vivo regional target occupancy (Occ) was obtained by pretreating with different doses of MP-10 (0. 05–2. 00mg/kg). Both in vitro binding assays and in vitro autoradiographic studies revealed a nanomolar binding affinity of [3H]TZ1964B to the rat striatum. The striatal binding of [3H]TZ1964B and [11C]TZ1964B was either displaced or blocked by MP-10 in rats and NHPs. Autoradiography and microPET imaging confirmed that the specific binding of the radioligand was found in the striatum but not in the cerebellum. Blocking studies also confirmed the suitability of the cerebellum as an appropriate reference region. The binding potentials (BPND) of [11C]TZ1964B in the NHP striatum that were calculated using either the Logan reference model (LoganREF, 3. 96±0. 17) or the simplified reference tissue model (SRTM, 4. 64±0. 47), with the cerebellum as the reference region, was high and had good reproducibility. The occupancy studies indicated a MP-10 dose of 0. 31±0. 09mg/kg (LoganREF)/0. 45±0. 17mg/kg (SRTM) occupies 50% striatal PDE10A binding sites. Studies in rats and NHPs demonstrated radiolabeled TZ1964B has a high binding affinity and good specificity for PDE10A, as well as favorable in vivo pharmacokinetic properties and binding profiles. Our data suggests that [11C]TZ1964B is a promising radioligand for in vivo imaging PDE10A in the brain of living subject.

AIIM Journal 2014 Journal Article

Vicinal support vector classifier using supervised kernel-based clustering

Xulei Yang
Aize Cao
Qing Song
Gerald Schaefer
Yi Su

Objective Support vector machines (SVMs) have drawn considerable attention due to their high generalisation ability and superior classification performance compared to other pattern recognition algorithms. However, the assumption that the learning data is identically generated from unknown probability distributions may limit the application of SVMs for real problems. In this paper, we propose a vicinal support vector classifier (VSVC) which is shown to be able to effectively handle practical applications where the learning data may originate from different probability distributions. Methods The proposed VSVC method utilises a set of new vicinal kernel functions which are constructed based on supervised clustering in the kernel-induced feature space. Our proposed approach comprises two steps. In the clustering step, a supervised kernel-based deterministic annealing (SKDA) clustering algorithm is employed to partition the training data into different soft vicinal areas of the feature space in order to construct the vicinal kernel functions. In the training step, the SVM technique is used to minimise the vicinal risk function under the constraints of the vicinal areas defined in the SKDA clustering step. Results Experimental results on both artificial and real medical datasets show our proposed VSVC achieves better classification accuracy and lower computational time compared to a standard SVM. For an artificial dataset constructed from non-separated data, the classification accuracy of VSVC is between 95. 5% and 96. 25% (using different cluster numbers) which compares favourably to the 94. 5% achieved by SVM. The VSVC training time is between 8. 75s and 17. 83s (for 2–8 clusters), considerable less than the 65. 0s required by SVM. On a real mammography dataset, the best classification accuracy of VSVC is 85. 7% and thus clearly outperforms a standard SVM which obtains an accuracy of only 82. 1%. A similar performance improvement is confirmed on two further real datasets, a breast cancer dataset (74. 01% vs. 72. 52%) and a heart dataset (84. 77% vs. 83. 81%), coupled with a reduction in terms of learning time (32. 07s vs. 92. 08s and 25. 00s vs. 53. 31s, respectively). Furthermore, the VSVC results in the number of support vectors being equal to the specified cluster number, and hence in a much sparser solution compared to a standard SVM. Conclusion Incorporating a supervised clustering algorithm into the SVM technique leads to a sparse but effective solution, while making the proposed VSVC adaptive to different probability distributions of the training data.