#17 Deep Learning for Personalized and Interpretable fMRI Analysis скачать в хорошем качестве

#17 Deep Learning for Personalized and Interpretable fMRI Analysis

Speaker: Yanting Yang (PhD Student at the University of British Columbia) Yanting Yang is a PhD student in the Department of Electrical and Computer Engineering at the University of British Columbia, advised by Dr. Xiaoxiao Li at the Trusted and Efficient AI (TEA) Lab. His research aims to improve the explainability, fairness, and efficiency of AI models, translating advanced machine learning algorithms into real-world healthcare applications. His work primarily involves developing novel deep learning architectures to analyze complex neuroimaging data, with a specific focus on functional Magnetic Resonance Imaging (fMRI). He investigates methods for personalized brain network modeling and cross-modal alignment to identify robust clinical biomarkers for neurological disorders. Deep learning offers significant potential for understanding neurological conditions, yet standard approaches often struggle with the trade-off between population-level consistency and individual variability, as well as the challenge of ensuring fairness. In this talk, Yang will explore two novel frameworks designed to address these limitations. First, Yang will introduce FunFormer, a method that uncovers individualized functional networks by incorporating learnable "prompt" tokens as dynamic cluster centroids. This approach captures both the stable group-level scaffold and the meaningful, personalized deviations that define individual brain function. He will demonstrate how these personalized networks serve as reliable neural fingerprints and improve prediction for neurodevelopmental disorders such as ASD and ADHD. Second, Yang will present NeuroLIP, a cross-modal alignment framework that integrates fMRI data with phenotypic text (e.g., diagnostic labels and demographics). Unlike global alignment methods, NeuroLIP utilizes localized tokens and text-conditioned attention to preserve fine-grained neuroanatomical interpretability. He will also discuss novel debiasing mechanisms—specifically negative gradient techniques and attention distance loss—that prevent models from exploiting sensitive attributes like age or sex, ensuring fair and clinically valid representations. Publications: https://arxiv.org/abs/2403.08203 https://arxiv.org/abs/2503.21964