IMDF-Net: Iterative U-Net With Multi-Kernel Dilated Convolution and Fusion Modules for Enhanced Retinal Vessel Segmentation

Abstract

In the early diagnosis of diabetic retinopathy, the morphological properties of blood vessels serve as an important reference for doctors to assess a patient's condition, facilitating scientific diagnostic and therapeutic interventions. However, vascular deformations, proliferation, and rupture caused by retinal diseases are often difficult to detect in the early stages. The assessment of retinal vessel morphology is subjective, time-consuming, and heavily dependent on the professional experience of the physician. Therefore, computer-aided diagnostic systems have gradually played a significant role in this field. Existing neural networks, particularly U-Net and its variants, have shown promising results in retinal vessel segmentation. However, due to the information loss caused by multiple pooling operations and the insufficient handling of local contextual features in skip connections, most segmentation methods still face challenges in accurately detecting microvessels. To address these limitations and assist medical staff in the early diagnosis of retinal diseases, we propose an iterative retinal vessel segmentation network with multi-dimensional attention and multi-scale feature fusion, named IMDF-Net. The network consists of a backbone network and an iterative refinement network. In the backbone network, we have designed a cascaded multi-kernel dilated convolution module and a multi-scale feature fusion module during the upsampling phase. These components expand the receptive field, effectively combine global information and local features, and propagate deep features to the shallow layers. Additionally, we have designed an iterative network to further capture missing information and correct erroneous segmentation results. Experimental results demonstrate that IMDF-Net outperforms several state-of-the-art methods on the DRIVE dataset, achieving the best performance across all evaluation metrics. On the CHASE_DB1 dataset, it achieves optimal performance in four metrics. It demonstrates its superiority in both overall performance and visual results, with a significant improvement in the segmentation of microvessels.