MRFANet: Massive retinopathy feature aggregation network for pixel-level diabetes-induced lesion detection from fundus images

Automatically segmenting diabetic retinopathy from fundus images through pixel-level lesion detection is of great importance for ophthalmologists to make accurate diagnosis. Although deep learning approaches based on convolutional neural network have already made great achievements in segmenting interested objects from images, they face challenges with retinopathy tasks due to the various sizes and the similarities between different lesions. To address this problem, this paper proposes a novel massive retinopathy feature aggregation network (MRFANet) for accurate multi-lesion segmentation of diabetic retinopathy. It can aggregate massive retinopathy features to make pixel-level predictions for different types of lesions in an end-to-end manner. We adopted successive dilated convolutions with residual links to realize hierarchical context feature extractions, and a multiscale pooling approach was applied to enrich semantic representations, then the massive features were aggregated to predict the final segmentations. Experiments were conducted on IDRiD and DDR datasets, and generalization performance was also evaluated, the results show that the proposed method is very competitive compared with the state-of-the-art approaches, especially for the soft exudate segmentation, of which the AUC scores are 0.7566 and 0.3286 on IDRiD and DDR datasets, respectively, achieving best performance in all the test scenarios.