Multi-Scale Feature Fusion Networks for Identifying Abnormal Brain Networks
Published in IEEE TNNLS (Under Review), 2025
Abstract
Existing brain network models analyze static and dynamic functional connectivity at a single scale, failing to capture hierarchical spatiotemporal dependencies. We propose MSD-Net, integrating group-guided contrastive pooling (static) and diffusion spatial attention (dynamic) via cross-view fusion, with anatomical-functional dual-scale tokenization and local-global fusion. The model is optimized by RevCon and entropy-based sparsity regularization, achieving state-of-the-art performance on three private TLE datasets and two public benchmarks (ABIDE-I, ADHD-200). This work was presented at ISICDM 2025 and the 2026 Chinese Intelligent Health & Bioinformatics Conference.
Recommended citation: Fang, Z., Chen, H.*, Huang, W.*, Li, R.*. (Under Review). Multi-Scale Feature Fusion Networks for Identifying Abnormal Brain Networks. IEEE Transactions on Neural Networks and Learning Systems.
Download Paper
