SegDRoWS: Segmentation of diabetic retinopathy lesions by a whole-stage multi-scale feature fusion network

Automatic segmentation of diabetic retinopathy (DR) lesions significantly aids ophthalmologists in diagnosis. The lesions often exhibit high similarity across classes, significant scale variances, tiny sizes and fuzzy edges, posing a formidable challenge for multi-class DR lesion segmentation. In this paper, a whole-stage multi-scale feature fusion network, termed SegDRoWS, is proposed to enhance the precision of DR segmentation. It consists of a three-stage encoder with intra-stage multi-scale feature fusion (IMFF), a detail-preserved inter-stage feature fusion (DIFF) block, an edge guidance branch (EGB) and a lightweight decoder. The IMFF encoder is introduced to explore intra-stage multi-scale features at granular level, utilizing different filter sizes to extract and fuse multi-scale features. Considering the importance of details for the segmentation of tiny lesions, the DIFF block is proposed to preserve details and play the role of inter-stage multi-scale feature fusion at the same time. To guide the model pay more attention on edge and detail information, the EGB is introduced. By combining the aforementioned elements, our SegDRoWS has the characteristics of “whole-stage multi-scale feature fusion”, as both intra- and inter-stage features are well explored. Our SegDRoWS achieves new state-of-the-art results on three public datasets with just 2.27M parameters, which is nearly 31 times fewer than the leading method, holding significant promise for clinical use.