DETF-Net: A Network for Retinal Vessel Segmentation Utilizing Detailed Feature Enhancement and Dynamic Temporal Fusion

The segmentation of retinal vessel images is a pivotal step in diagnosing various ophthalmic and systemic diseases. Among deep learning techniques, UNet has been extensively utilized for its capability to deliver remarkable segmentation results. Nonetheless, significant challenges persist, particularly the loss of detail and spatial resolution caused by downsampling operations in convolutional and pooling layers. This drawback often results in subpar segmentation of small targets and intricate boundaries. Furthermore, achieving a balance between capturing global context and preserving local detail remains challenging, thereby limiting the segmentation performance on multi-scale targets. To tackle these challenges, this study proposes the Detail-Enhanced Temporal Fusion Network (DETF-Net), which introduces two essential modules: (1) the Detail Feature Enhancement Module (DFEM), designed to strengthen the representation of complex boundary features through the integration of median pooling, spatial attention, and mixed depthwise convolution; and (2) the Dynamic Temporal Fusion Module (DTFM), which combines Multi-scale Feature Extraction (MFE) and the Temporal Fusion Attention Mechanism (TFAM). The MFE module improves robustness across varying vessel sizes and shapes, while the TFAM dynamically adjusts feature importance and effectively captures subtle changes in vessel structure. The effectiveness of DETF-Net was evaluated on three benchmark datasets: DRIVE, CHASE_DB1, and STARE. The proposed network achieved high accuracy scores of 0.9811, 0.9875, and 0.9876, respectively, alongside specificity values of 0.9811, 0.9870, and 0.9875. Comparative experiments demonstrated that DETF-Net outperforms current state-of-the-art models, showcasing its superior segmentation performance. This research presents innovative approaches to address existing limitations in retinal vessel image segmentation, thereby advancing diagnostic accuracy for ophthalmic diseases.