Dynamic multi-scale fusion and dual-process convolution for improved retinal lesion segmentation

Accurate segmentation of diabetic retinopathy lesions from fundus images remains a significant challenge due to their diverse scales and subtle visual characteristics. To tackle this issue, a new architecture called DP-UNet is introduced, combining a contemporary ConvNeXt backbone with a U-Net framework. The framework is enhanced through the synergistic incorporation of a Dynamic Multi-Scale Fusion (DMSF) module for adaptive context aggregation, the Dual-Process Convolution (DPConv) block for feature enhancement, and a wavelet-based dual-path downsampling mechanism to preserve high-frequency edge details. Comprehensive evaluations on the IDRiD (Porwal et al., Med Image Anal 59:101561, 2020) and DDR (Li et al., Inf Sci 501:511–522, 2019) datasets demonstrate that the proposed method achieves state-of-the-art performance. It attains a mean PR-AUC of 63.32% on IDRiD and 58.78% on DDR, with superior results specifically in hemorrhage segmentation. Ablation studies confirm the individual contributions of each component, particularly in improving the segmentation of microaneurysms, a critical yet challenging lesion type. The architecture effectively reconciles the extraction of global semantic information with the preservation of local fine-grained details, advancing the task from pixel-wise classification to multi-scale feature synergy. While the model exhibits strong overall performance, qualitative analysis identifies minor boundary inaccuracies, highlighting an avenue for future refinement. This work provides a robust and generalizable solution for automated lesion segmentation, offering considerable potential to assist in clinical diagnosis and management.