Microaneurysm Segmentation Method for Diabetic Retinopathy Fundus Lesions Based on the Multi-Scale Feature Subtraction Fusion Network

Detecting and segmenting microaneurysms can help doctors diagnose the condition and formulate subsequent treatment plans. A multi-scale feature subtraction fusion network is proposed in this paper. It includes two modules: the multi-scale feature compensation module and the subtraction fusion module. In the multi-scale feature compensation module, the features between adjacent levels of the network are fused. Considering that simply concatenating features may lead to feature redundancy, a subtraction fusion module is designed. To enable the neural network to extract more detailed information, a branch is introduced. A wavelet attention enhancement module is designed to transform the channel attention of frequency coefficients extracted by wavelet transform. The proposed method can help the network learn feature diversity better, and hence can improve segmentation performance. Experimental results show that, as compared to the existing methods, the proposed method can achieve better performance with Dice coefficients of 0.4481, 0.4860, and 0.3561 on the IDRID, E-Ophtha, and DDR datasets, respectively.