Multi-Scale Deep Supervised Attention Network for Red Lesion Segmentation

Abstract

Diabetic Retinopathy (DR) is one of the leading causes of blindness among the diabetic population. Early diagnosis of DR can prevent visual impairment and leads to the motivation of designing an automatic DR detection model. Non-proliferative Diabetic Retinopathy (NPDR) is detected by the formation of Red Lesions - MicroAneurysms and Hemorrhages. MicroAneurysms are minute in structure and need special care to be located. This paper targets the automatic detection of DR at the preliminary stage by implementing a modified Full-scale Deeply Supervised Attention Network (FuDSA-Net). The architecture encompasses a multi-scale feature-based attention module along with deep supervision to help achieve high-quality segmentation output. Experimental results suggest that the model with focal Tversky loss outperforms state-of-the-art architectures.