Visualizing sex differences in deep learning models for neuroimaging using neuroanatomically guided salient source separation

Motivation: This work aims at utilizing and interpreting the associative patterns learned by DL models with the goal of aiding data-driven biomarker development for brain disorders. Goal(s): Automatically source separation of salient brain areas from the high-dimensional predictive encodings of deep learning models for structural neuroimaging data. Approach: We compute active subspaces in the saliency space of a deep learning network trained using sMRI maps, followed by a spatially-constrained independent component analysis (scICA) step. Results: Our framework is able to compute multiple salient brain sources that characterize the sex differences encoded by a deep learning model, along with high prediction accuracy. Impact: This work aims at utilizing and interpreting the associative patterns learned by DL models with the goal of aiding data-driven biomarker development for brain characeristics and disorders.