DDFU-Net: A Deep Decoder-Focused U-Net Model for Retinal Lesion Segmentation

Early detection of retinal lesions helps to avoid visual loss or blindness. The main lesions associated with eye diseases include soft exudates, hard exudates, microaneurysms, and hemorrhages. However, the segmentation of these four kinds of lesions is difficult and time-consuming due to their uncertainty in size, contrast, and high inter-class similarity. To address these issues, this study presents Deep Decoder-Focused U-Net (DDFU-Net), an asymmetric dense U-Net model for automatic and accurate multi-lesion segmentation using fundus images. Our approach simultaneously segments all four kinds of retinal lesions after proving that multi-task learning yields better results than single-task learning. DDFU-Net incorporates an asymmetric design with five dense blocks in the encoder and seven dense blocks in the decoder. This design enhances feature extraction while ensuring a more refined reconstruction of lesion boundaries, particularly for small and complex structures. By allocating more layers to the decoder, the model improves segmentation accuracy by gradually restoring spatial details lost during down-sampling, mitigating over-compression, and enhancing fine-grained feature preservation. Comprehensive experiments on IDRiD and DDR datasets well demonstrate the superiority of our approach, which outperforms state-of-the-art segmentation methods. Specifically, DDFU-Net achieved a mean Area Under the Precision-Recall Curve of 54.86%, a mean Intersection Over Union of 36.96%, and mean Dice scores of 52.24% on the DDR test set. On the IDRiD test set, it achieved 66.69%, 57.31%, and 69.93%, respectively. The asymmetric structure outperforms traditional symmetric U-Nets by capturing more detailed features during encoding while reducing complexity during decoding. The proposed method can be useful to aid in the diagnosis of eye diseases, reducing the workload of specialists and improving the attention to patients.