CARNet: Cascade attentive RefineNet for multi-lesion segmentation of diabetic retinopathy images

Abstract Diabetic retinopathy is the leading cause of blindness in working population. Lesion segmentation from fundus images helps ophthalmologists accurately diagnose and grade diabetic retinopathy. However, task lesion full challenges due to complex structure, various sizes interclass similarity with other tissues. To address issue, this paper proposes a cascade attentive RefineNet (CARNet) for automatic accurate multi-lesion It can make use fine local details coarse global information image. CARNet composed image encoder, encoder attention refinement decoder. We take whole patch as dual input, feed them ResNet50 ResNet101, respectively, downsampling extract features. The high-level decoder uses mechanism integrate same-level features two encoders output low-level module multiscale fusion, which focus model on area generate predictions. evaluated performance proposed IDRiD, E-ophtha DDR data sets. Extensive comparison experiments ablation studies sets demonstrate framework outperforms state-of-the-art approaches has better accuracy robustness. not only overcomes interference similar tissues noises achieve segmentation, but also preserves contour shape small lesions without overloading GPU memory usage.