Phase and structure based input enhancement for retinal vessel segmentation

This paper presents a study of retinal vessel segmentation that combines traditional image processing methods with deep learning. Two edge detection techniques, Phase Stretch Transform (PST) and B-COSFIRE filters. Both were applied as preprocessing steps to enhance vessel structures before segmentation using a U-net. PST parameters were optimized via a parallel genetic algorithm, and the resulting vessel maps replaced one channel in the RGB color space of the original fundus images. The method was evaluated on five public datasets (DRIVE, CHASE DB1, HRF, TREND, and FIVES) using standard segmentation metrics. Results show improved performance on DRIVE, CHASE and HRF datasets, and comparable or slightly improved performance on other datasets, indicating that spectral and structural enhancement can be complementary to deep learning methods without increasing computational complexity.