Interpretable Deep Learning for Analyzing Brain Development Heterogeneity with Personalized Functional Networks Across Multi-Site Data

ABSTRACT Intrinsic functional connectivity magnetic resonance imaging (fcMRI) is a powerful tool for understanding brain function and development. Despite significant progress in modeling personalized functional networks (PFNs) and leveraging machine learning (ML) to predict brain-behavior relationships, challenges persist, including the heterogeneity of brain development, the difficulty in characterizing diverse PFN patterns, site effects in multi- site datasets, and the limited interpretability of deep learning (DL) models. To address these challenges, this project aims to develop, validate, and disseminate tools for computing and characterizing heterogeneous PFNs, harmonizing multi-site fcMRI data, and building interpretable prediction models. These models will capture PFN patterns linked to variations in brain development and psychopathology in youth. Specifically, we will develop a novel self-supervised DL method to identify distinct biotypes of individual variation in fcMRI data and compute biotype-specific PFNs using a mixture-of-experts approach. We will also develop a robust self- supervised DL method to harmonize multi-site fcMRI data and compute PFNs, employing a test-time adaptation (TTA) strategy. Additionally, we will develop an interpretable DL method to characterize and predict cognition and psychopathology based on PFNs, leveraging a generalized additive model (GAM) framework with a prototype learning mechanism. The tools will be developed and validated using data from the Adolescent Brain Cognitive Development (ABCD) Study, the Philadelphia Neurodevelopmental Cohort (PNC), the Healthy Brain Network (HBN), and the Lifespan Human Connectome Project in Development (HCP-D). The algorithms will be released as a user-friendly toolbox with source code and standalone programs on GitHub and DockerHub. This project will revolutionize our understanding of brain development, enhance our understanding of individual brain development, and pave the way for more effective, personalized treatments in diverse clinical settings.