CMAC-Net: Cascade Multi-Scale Attention Convolution Network for diabetic retinopathy lesion segmentation

Automatic diabetic retinopathy (DR) lesions is of great importance to assist ophthalmologists in making accurate diagnoses. Currently, the accurate segmentation of fundus lesions is challenged by the large differences in the shape and size of the lesions, and their concentrated locations of different lesions. Many deep learning methods based on convolutional neural networks (CNN) have been proposed to handle the task of segmentation of fundus lesion. However, these methods still have two limitations: (1) lack of backbone networks targeting the characteristics of fundus images, and limited ability to extract lesions of different scales; (2) with semantic conflicts when fusing multi-level information from nonadjacent layers. To this end, we propose a Cascade Multi-Scale Attention Convolution Network (CMAC-Net) for diabetic retinopathy lesion segmentation. In particular, Mobile Dynamic Attention Convolution (MDAC) is adopted as backbone network integrating with Cascade Progressive Context Fusion (CPCF) module in a U-Net framework. CMAC-Net has the following advantages: (1) MDAC preserves high-resolution features, leading to high segmentation accuracy of minor lesions; (2) MDAC uses large kernel convolution to separate the large and small lesions at the same time; (3) CPCF aims to fuse the semantic information from different levels, avoiding the inconsistent parts. The experimental results on three datasets show that our model outperforms the state-of-the-art, in terms of several metrics.