Incomplete Multimodal Alzheimer's Disease Classification via Bidirectional GAN and Spectral Graph Learning

Bo Cheng, Zhi Yang, Haitao Gan, Zhongwei Huang, Ran Zhou, Ming Shi, Lei Cheng

Abstract: The early diagnosis of Alzheimer's disease (AD) relies on the joint analysis of multimodal neuroimaging data, but the modality missing problem severely limits diagnostic performance. In this paper, we propose an innovative framework integrating bidirectional generative adversarial networks with adaptive spectral graph learning: (1) We design a bidirectional generative adversarial network (BiGAN) model, which realizes the bidirectional generation of structural connectivity (SC) and functional connectivity (FC) networks through cyclic consistency constraints and symmetric positive definite (SPD) manifold optimization, solving the modality missing problem; (2) We propose a Dual-Branch Attentive Spectral Wavelet-GCN (DASW-GCN) model based on spectral graph learning, which adaptively fuses multimodal brain graph features by utilizing spectral wavelet transform and cross-modal attention mechanisms. Experiments were conducted on the ADNI dataset. The proposed method shows significant improvement compared to baseline models in the AD classification task. Geometric score (GS) analysis indicates that the topological fidelity of the generated FC networks is superior to that of traditional GAN models. This study provides a new idea for the early diagnosis of AD in scenarios with multimodal missing data.

External IDs:dblp:conf/icic/ChengYGHZSC25