LezioSeg: Multi-Scale Attention Affine-Based CNN for Segmenting Diabetic Retinopathy Lesions in Images

Diagnosing some eye pathologies, such as diabetic retinopathy (DR), depends on accurately detecting retinal eye lesions. Automatic lesion-segmentation methods based on deep learning involve heavy-weight models and have yet to produce the desired quality of results. This paper presents a new deep learning method for segmenting the four types of DR lesions found in eye fundus images. The method, called LezioSeg, is based on multi-scale modules and gated skip connections. It has three components: (1) Two multi-scale modules, the first is atrous spatial pyramid pooling (ASPP), which is inserted at the neck of the network, while the second is added at the end of the decoder to improve the fundus image feature extraction; (2) ImageNet MobileNet encoder; and (3) gated skip connection (GSC) mechanism for improving the ability to obtain information about retinal eye lesions. Experiments using affine-based transformation techniques showed that this architecture improved the performance in lesion segmentation on the well-known IDRiD and E-ophtha datasets. Considering the AUPR standard metric, for the IDRiD dataset, we obtained 81% for soft exudates, 86% for hard exudates, 69% for hemorrhages, and 40% for microaneurysms. For the E-ophtha dataset, we achieved an AUPR of 63% for hard exudates and 37.5% for microaneurysms. These results show that our model with affine-based augmentation achieved competitive results compared to several cutting-edge techniques, but with a model with much fewer parameters.