Multimodal Fusion Network with Information Bottleneck Mamba and Intervention Enhancement for Retinal Disease...

Multimodal fundus disease diagnosis holds significant clinical value. Color fundus photography (CFP) captures superficial retinal vasculature and lesion distribution, while optical coherence tomography (OCT) reveals the microstructure of deep retinal tissues, providing complementary pathological information. However, existing multimodal approaches still suffer from redundant non-diagnostic features, spurious correlations, and insufficient fusion granularity, hindering accurate identification of complex retinal pathologies. To address these challenges, we propose IMIE-Net, a multimodal fusion network based on the Mamba information bottleneck and intervention-enhanced mechanisms. First, a multimodal feature extraction filter uses Mamba networks with different scanning strategies to extract modality-specific features from CFP and OCT, followed by an information bottleneck that removes irrelevant features to retain disease-relevant features. Second, an interference feature intensifier employs an interference-aware dynamic recognition mechanism to strengthen relevant feature dimensions while suppressing spurious correlations, enhancing discriminative power. Finally, a cross-modal fusion dynamic decision classifier integrates multimodal features through explicit and implicit dual-path fusion and performs dynamic decision optimization to improve diagnostic robustness. Extensive experiments on multiple fundus datasets have shown that IMIE-Net can achieve accurate and reliable diagnosis of fundus diseases.