A Robust Graph-based Method with Junction Detection and Angle Pruning for Longest Continuous Non-branching Vessel Segmentation in OCTA Images

Abstract

Optical coherence tomography angiography (OCTA) is a powerful imaging technique for non- invasively visualizing retinal blood flow at high resolution. Accurate vessel segmentation from OCTA images is essential for diagnosing and monitoring retinal diseases. However, segmentation remains challenging due to the complexity of the vessel network, including sharp turns, varying vessel widths, and frequent junctions. Additionally, denoising OCTA images without losing fine vessel structures further complicates the process. This study proposes an enhanced graph traversal method for OCTA vessel segmentation. Incorporating angular-threshold-based pruning and improved junction handling to address common challenges with a comprehensive preprocessing pipeline to denoise the OCTA image while preserving vessel integrity. It also gives a comparative analysis against a baseline graph traversal technique, which extracts all vessel paths without pruning or junction refinement. This research aims to enhance the accuracy of extraction of longest biologically realistic continuous vessel segments from an OCTA image. To evaluate our method, we use a dataset of five OCTA images each comprising of approximately 175 vessel segments and 60 longest vessel strains. Evaluation metrics include false positive rates and qualitative visual comparisons. Visual analysis demonstrates that our pruning technique significantly improves segmentation quality, producing smoother, continuous and biologically valid vessels while reducing spurious branches. Our results yield an F1 score of 0.8488, showing a marked improvement over the baseline model.