SepFE: Separable Fusion Enhanced Network for Retinal Vessel Segmentation

The accurate and automatic segmentation of retinal vessels from fundus images is critical for the early diagnosis prevention many eye diseases, such as diabetic retinopathy (DR). Existing vessel approaches based on convolutional neural networks (CNNs) have achieved remarkable effectiveness. Here, we extend a model with low complexity high performance U-Net, which one most popular architectures. In view excellent work depth-wise separable convolution, introduce it to replace standard layer. proposed reduced by decreasing number parameters calculations required model. To ensure while lowering redundant parameters, integrate pre-trained MobileNet V2 into encoder. Then, feature fusion residual module (FFRM) designed facilitate complementary strengths enhancing effective between adjacent levels, alleviates extraneous clutter introduced direct fusion. Finally, provide detailed comparisons SepFE U-Net in three image mainstream datasets (DRIVE, STARE, CHASEDB1). results show that only 3% Flops are 8% better obtained. superiority further demonstrated through other advanced methods.