D 3 -Mamba: A Dual-Stream Mamba Network for Detecting Optic Disc Edema from Color Fundus Photographs

Accurate differentiation between optic disc edema (ODE) and other optic disc anomalies, such as pseudopapilledema (PPE), is vital yet challenging in neuro-ophthalmology due to their highly overlapping visual features on 2D fundus images. This paper proposes the D 3 -Mamba (Dual-branch, Dual-difference, Dual-domain Mamba), a novel hybrid architecture designed to capture fine-grained discriminative features for robust classification. First, we introduce the Hierarchical Mamba Serialization Strategies (HMSS) to tailor scanning mechanisms to different feature resolutions. Featuring an adaptive spiral scan mechanism, this strategy is designed to explicitly model the concentric structure of the optic disc, mimicking the focused visual path of clinicians. Additionally, a History-aware Dual-Difference Interaction (HDDI) module is incorporated to facilitate feature interaction between the CNN and Mamba branches via a recursive history update mechanism, ensuring that global context and local details are effectively fused. Finally, the Dual-Domain Aggregation (DDA) network integrates spatial features with frequency-domain analysis to compensate for the information loss inherent in spatial downsampling, thereby enriching the representation of complex pathological patterns. The model was evaluated on the Papilledema2018 and the Eye Disease Image Dataset. Experimental results demonstrate that D 3 -Mamba achieves robust diagnostic performance compared to mainstream deep learning architectures. Notably, the model exhibited high sensitivity for detecting ODE across both datasets, highlighting its potential utility in reducing clinical misdiagnosis risks. This research presents a promising approach for automated primary care screening, offering a fair balance between diagnostic reliability and computational efficiency.