Deep learning has disrupted nearly every major field of study from computer vision to genomics. The unparalleled success of these models has, in many cases, been fueled by an explosion of data. Millions of labeled images, thousands of annotated ICU admissions, and hundreds of hours of transcribed speech are common standards in the literature. Clinical neuroscience is a notable holdout to this trend. It is a field of unavoidably small datasets, massive patient variability, and complex (largely unknown) phenomena. My lab tackles these challenges across a spectrum of projects, from answering foundational neuroscientific questions to translational applications of neuroimaging data to exploratory directions for probing neural circuitry. One of our key strategies is to integrate a priori information about the brain and biology into the model design.

This talk will highlight two ongoing projects that epitomize this strategy. First, I will showcase an end-to-end deep learning framework that fuses neuroimaging, genetic, and phenotypic data, while maintaining interpretability of the extracted biomarkers. We use a learnable dropout layer to extract a sparse subset of predictive imaging features and a biologically informed deep network architecture for whole-genome analysis. Specifically, the network uses hierarchical graph convolution that mimic the organization of a well-established gene ontology to track the convergence of genetic risk across biological pathways. Second, I will present a deep-generative hybrid model for epileptic seizure detection from scalp EEG. The latent variables in this model capture the spatiotemporal spread of a seizure; they are complemented by a nonparametric likelihood based on convolutional neural networks. I will also highlight our current end-to-end extensions of this work focused on seizure onset localization. Finally, I will conclude with exciting future directions for our work across the foundational, translational, and exploratory axes.

Roughly 40% of the population will be diagnosed with some form of cancer in their lifetime. In a large majority of these cases, a definitive cancer diagnosis is only possible via histopathologic confirmation on a tissue slide. With the increasing popularity of the digitization of pathology slides, a wealth of new untapped data is now regularly being created.

Computational analysis of these routinely captured H&E slides is facilitating the creation of diagnostic tools for tasks such as disease identification and grading. Further, by identifying patterns of disease presentation across large cohorts of retrospectively analyzed patients, new insights for predicting prognosis and therapy response are possible [1,2]. Such biomarkers, derived from inexpensive histology slides, stand to improve the standard of care for all patient populations, especially where expensive genomic testing may not be readily available. Moreover, since numerous other diseases and disorders, such as oncoming clinical heart failure [3], are similarly diagnosed via pathology slides, those patients also stand to benefit from these same technological advances in the digital pathology space.

This talk will discuss our research aimed towards reaching the goal of precision medicine, wherein patients receive optimized treatment based on historical evidence. The talk discusses how the applications of deep learning in this domain are significantly improving the efficiency and robustness of these models [4]. Numerous challenges remain, though, especially in the context of quality control and annotation gathering. This talk further introduces the audience to open-source tools being developed and deployed to meet these pressing needs, including quality control (histoqc.com [5]), annotation (quickannotator.com), labeling (patchsorter.com), validation (cohortfinder.com).

This presentation will identify the gap between AI accuracy and making good clinical decisions. I will present a study where we develop an ethical framework for clinical decision-making that can help clinicians meet medicolegal and ethical standards when using AI that does not rely on explainability, nor perfect accuracy of the model.