CFAM-Net: Cascade Network Based on Full-Resolution and Attention Mechanism for Vessel Segmentation

The morphology of retinal vessels in fundus images is essential for diagnosing ophthalmic diseases like diabetic retinopathy, glaucoma, and conjunctivitis. Traditional analysis methods are often time-consuming, labor-intensive, and prone to errors due to image noise and low contrast, resulting in inaccurate diagnoses. To overcome these limitations, this paper introduces a cascade network using full-resolution representation and attention mechanisms for precise segmentation of fine vessels. Multi-resolution convolutional layers are employed to reduce spatial information loss and retain detailed vessel structures. A feature aggregation attention module is integrated to improve multi-scale vessel information extraction. Finally, a refined multi-semantic fusion module enhances vessel connectivity and supervises the segmentation process. The method achieved outstanding performance, with accuracy, sensitivity, and F1 scores of 0.9707, 0.8312, and 0.8319 on the DRIVE dataset, surpassing existing approaches and offering promising clinical relevance.