Learning robust and general representations of the brain using large heterogeneous multi-study neuroimaging data Vishnu Bashyam

Format:

Book

Thesis/Dissertation

Author/Creator:

Bashyam, Vishnu, author.

Contributor:

University of Pennsylvania. Bioengineering., degree granting institution.

Language:

English

Subjects (All):

Biomedical engineering.

Bioinformatics.

Medical imaging.

0800.

0541.

0574.

0715.

Local Subjects:

Biomedical engineering.

Bioinformatics.

Medical imaging.

0800.

0541.

0574.

0715.

Genre:

Academic theses

Physical Description:

1 online resource (119 pages)

Contained In:

Dissertations Abstracts International 87-12B

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Deep learning has enabled powerful neuroimaging models, yet most remain brittle when deployed across scanners, acquisition protocols, and study populations, and they often require task-specific labeled datasets that are too small for reliable training. This dissertation addresses these two barriers by learning robust and general representations of brain structure from large, heterogeneous, multi-study MRI data. Using tens of thousands of T1-weighted scans aggregated across diverse cohorts, we develop methods that (1) reduce site- and scanner-driven variation that confounds downstream prediction and (2) leverage supervision from multiple tasks to build transferable feature representations that improve sample efficiency for new neuroimaging tasks.First, we investigate image-level harmonization using deep generative models to map scans from disparate sites into a common reference domain while preserving subject-specific anatomy. Using unpaired multi-site data, we show that GAN-based harmonization can substantially improve out-of-sample generalization for brain age prediction relative to unharmonized and histogram-matched baselines, demonstrating that explicitly modeling acquisition-related variation can enhance cross-site portability of predictors.Second, we develop a brain-specialized pretrained model via large-scale brain age prediction, demonstrating that representations learned from age supervision transfer more effectively to downstream disease classification than conventional ImageNet initialization. We further show that the tightest age-fitting models are not necessarily the most clinically informative and that moderately regularized models yield brain-age deviations that better differentiate disease groups.Third, motivated by the success of single-task pretraining on brain age, we generalize this approach to learn from multiple clinically relevant tasks simultaneously. We introduce a framework for learning general-purpose brain representations from heterogeneous labels spanning cognition, risk factors, genetics, biomarkers, and diagnosis. By structuring knowledge transfer across tasks through sequential training, we show that multi-task pretraining further improves generalization beyond single-task approaches, particularly in low-data regimes and for entirely held-out prediction targets.Finally, we develop a fast, robust deep learning brain MRI segmentation pipeline validated at scale, enabling rapid extraction of anatomically meaningful features for broad downstream use.Collectively, this work establishes practical approaches for developing robust, generalizable deep learning models for neuroimaging. All models and tools are made publicly available to facilitate broader adoption

Notes:

Source: Dissertations Abstracts International, Volume: 87-12, Section: B.

Advisors: Davatzikos, Christos Committee members: Nasrallah, Ilya M.; Shou, Haochang; Shen, Li

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247980131

Access Restriction:

Restricted for use by site license