MAINet: Multi-scale attention interaction network for diabetic retinopathy lesion segmentation

Diabetic retinopathy ranks among the leading causes of blindness worldwide, severely impacting the quality of individual life. Therefore, early screening and timely diagnosis are of great importance in controlling disease progression and alleviating patient suffering. However, the complex structures, varying scales, scattered distribution and blurred edges of DR lesions pose significant challenges for automated and accurate segmentation. To address the issues of inconsistent scales and insufficient information interaction between adjacent encoding features, this paper introduces a multi-scale attention interaction network named MAINet for concurrently segmenting four types of diabetic retinopathy lesions. The VGG16 backbone is initially employed to extract features from retinal images. Then, the global-local attention module utilizes a parallel global-local architecture to capture global channel attention features, global spatial attention features and local detail features, thereby providing enriched contextual representations of the lesions. Moreover, the cross-attention interaction module performs a cross scale fusion of global-local attention features from different levels using a weight-sharing mechanism, which learns high-level semantic features and fine-grained detail features while mitigating information loss. Comprehensive comparative and ablation studies were performed on the IDRiD, DDR, and FGADR datasets. The experimental results indicate that the proposed method surpasses current leading models in segmentation performance, with significant improvements particularly in microaneurysms and soft exudates.