SDC-Net: Enhancing Multiclass Diabetic Retinopathy Segmentation by Shallow–Deep Collaboration Networks With Wavelet-Guided Attention Mechanism

Purpose: Diabetic retinopathy (DR) is a leading cause of blindness. Fundus lesions are key clinical signs of DR, and their accurate segmentation is crucial for screening, grading, and monitoring the disease. However, segmenting different lesions simultaneously is challenging owing to their varying shapes, sizes, and appearances. This work aimed to segment four types of DR lesions simultaneously. Methods: We propose a shallow-deep collaboration network with a wavelet-guided attention mechanism for simultaneous segmentation of four DR-related lesions. Our end-to-end framework integrates shallow and deep networks to enhance multiscale feature extraction. The deep network uses a wavelet-based attention mechanism to fuse multiscale context representations. Additionally, a super-resolution auxiliary task is introduced to improve training accuracy. Results: Extensive experiments are conducted on the IDRiD, DDR, and FGADR public datasets for evaluation. The average Dice scores of SDC-Net on the three datasets are 60.27, 39.98, and 42.57, with average intersection over union scores of 44.53, 25.71, and 28.58, and average area under the receiver operating characteristic curve values of 68.75, 57.37, and 64.21. Conclusions: The dual-branch design in the proposed framework enables better capture of multiscale features and improves segmentation. The dual wavelet attention module of the deep branch can enhance the extraction of detailed lesion features, and the introduced super-resolution task further improves the accuracy and robustness. Translational Relevance: The SDC-Net framework has the potential to enhance the clinical diagnosis of DR by offering more precise segmentation of multiple types of fundus lesions, thereby aiding in early detection and management of the disease.