MicroSeg: Multi-scale fusion learning for microaneurysms segmentation

Microaneurysms (MAs) are the initial clinical signs of diabetic retinopathy (DR), which is crucial for the early prevention and treatment of DR. However, detecting MAs in fundus images is challenging due to their presence as small, blood-colored dots amidst complex noise. Traditional single-stage Convolutional Neural Networks (CNN) struggle to extract sufficient contextual information for effective detection. To overcome this, we introduce MicroSeg, a multi-scale fusion learning model with a U-shaped structure. Utilizing the Swin Transformer as the backbone, we design two innovative modules: the Multi-Scale Pyramid Fusion (MSPF) and the Weight-sharing Parallel Dilated Convolutions (WPDC), which reconstruct skip connections to enhance cross-scale complementarity and compensate for underlying information gaps. Additionally, we propose the High Pass Overlay (HPO) preprocessing method to augment MAs feature representation. Our model demonstrates superior performance, achieving Free-response Receiver Operating Characteristic (FROC) curve scores of 0.535 on the E-Ophtha-MA dataset and 0.259 on the Dataset for Diabetic Retinopathy (DDR), outperforming comparable state-of-the-art (SOTA) end-to-end algorithms.