Retinal artery/vein classification by multi-channel multi-scale fusion network

The automatic artery/vein (A/V) classification in retinal fundus images plays a significant role in detecting vascular abnormalities and could speed up the diagnosis of various systemic diseases. Deep-learning methods have been extensively employed in this task. However, due to the lack of annotated data and the serious data imbalance, the performance of the existing methods is constricted. To address these limitations, we propose a novel multi-channel multi-scale fusion network (MMF-Net) that employs the enhancement of vessel structural information to constrain the A/V classification. First, the newly designed multi-channel (MM) module could extract the vessel structure from the original fundus image by the frequency filters, increasing the proportion of blood vessel pixels and reducing the influence caused by the background pixels. Second, the MMF-Net introduces a multi-scale transformation (MT) module, which could efficiently extract the information from the multi-channel feature representations. Third, the MMF-Net utilizes a multi-feature fusion (MF) module to improve the robustness of A/V classification by splitting and reorganizing the pixel feature from different scales. We validate our results on several public benchmark datasets. The experimental results show that the proposed method could achieve the best result compared with the existing state-of-the-art methods, which demonstrate the superior performance of the MMF-Net. The highly optimized Python implementations of our method is released at: https://github.com/chenchouyu/MMF_Net.