A frequency-driven method for retinal vessel segmentation

Retinal vessel segmentation is essential for diagnosing and monitoring ocular and systemic diseases. However, most existing deep learning methods operate predominantly in the spatial domain, limiting their ability to capture fine vessel structures and maintain topological continuity, especially for thin and low-contrast vessels. Existing frequency-aware approaches mostly rely on fixed transforms or simple feature concatenation, lacking task-adaptive spectral modeling. We developed a frequency-driven framework to address these limitations.

We implemented FRetinalNet, a frequency-driven encoder-decoder architecture. The encoder employs a Frequency Feature Tuning Block (FFTBlock) that parameterizes convolutional kernels in the spectral domain and adaptively modulates them with contextual cues via inverse FFT. Spatial and frequency features from parallel branches are fused using a Frequency Aware Fusion Block (FaFBlock) with cross-modality attention. The decoder includes dual-resolution branches for fine boundary recovery and global structural coherence, along with an auxiliary decoder for complementary supervision. Training proceeds in two stages: Stage 1 pretrains the encoder with original and Dual-Tree Complex Wavelet Transform inputs using consistency loss; Stage 2 transfers weights and fine-tunes under supervised objectives. We evaluated on DRIVE, STARE, and OCTA-500 using Dice, sensitivity, specificity, accuracy, and AUC.

FRetinalNet achieved the highest Dice scores on DRIVE (0.8578) and STARE (0.8696) compared with state-of-the-art methods. On OCTA-500, it obtained the best Dice for arteries (0.8864) and veins (0.9055) on the 6mm subset, and the best arterial (0.9169) and venous (0.9101) Dice on the 3mm subset. Ablation studies confirmed that two-stage training improves Dice from 0.8390 to 0.8578, and that all proposed components provide complementary benefits.

Explicit integration of learnable spectral parameterization and cross-domain feature fusion improves delineation of fine structures and topological continuity. The two-stage training pipeline provides robust frequency-aware initialization across fundus and OCTA modalities.