Author ORCID Identifier

Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Sergey Plis

Second Advisor

Rajshekhar Sunderraman

Third Advisor

Robyn Miller

Fourth Advisor

Vince Calhoun


Brain dynamics are highly complex and yet hold the key to understanding brain function and dysfunction. The dynamics captured by resting-state functional magnetic resonance imaging data are noisy, high-dimensional, and not readily interpretable. The typical approach of reducing this data to low-dimensional features and focusing on the most predictive features comes with strong assumptions and can miss essential aspects of the underlying dynamics. In contrast, introspection of discriminatively trained deep learning models may uncover disorder-relevant elements of the signal at the level of individual time points and spatial locations. Nevertheless, the difficulty of reliable training on high-dimensional but small-sample datasets and the unclear relevance of the resulting predictive markers prevent the widespread use of deep learning in functional neuroimaging. In this dissertation, we address these challenges by proposing a deep learning framework to learn from high-dimensional dynamical data while maintaining stable, ecologically valid interpretations. The developed model is pre-trainable and alleviates the need to collect an enormous amount of neuroimaging samples to achieve optimal training.

We also provide a quantitative validation module, Retain and Retrain (RAR), that can objectively verify the higher predictability of the dynamics learned by the model. Results successfully demonstrate that the proposed framework enables learning the fMRI dynamics directly from small data and capturing compact, stable interpretations of features predictive of function and dysfunction. We also comprehensively reviewed deep interpretability literature in the neuroimaging domain. Our analysis reveals the ongoing trend of interpretability practices in neuroimaging studies and identifies the gaps that should be addressed for effective human-machine collaboration in this domain.

This dissertation also proposed a post hoc interpretability method, Geometrically Guided Integrated Gradients (GGIG), that leverages geometric properties of the functional space as learned by a deep learning model. With extensive experiments and quantitative validation on MNIST and ImageNet datasets, we demonstrate that GGIG outperforms integrated gradients (IG), which is considered to be a popular interpretability method in the literature. As GGIG is able to identify the contours of the discriminative regions in the input space, GGIG may be useful in various medical imaging tasks where fine-grained localization as an explanation is beneficial.


File Upload Confirmation