LS-CF filter boosts vessel segmentation accuracy by 10% across 5 datasets

A team of researchers from multiple institutions (including Erick O. Rodrigues, Lucas O. Rodrigues, and Panos Liatsis) introduced the Local-Sensitive Connectivity Filter (LS-CF), a novel unsupervised post-processing technique designed to improve vessel segmentation in medical imaging. Traditional filters like Frangi, Hessian, and vesselness often produce broken vessel responses due to intensity inhomogeneities or noise. LS-CF addresses this by computing pixel-level vessel continuity while applying a local tolerance heuristic that intelligently reconnects fragmented vessels. The method requires no labeled training data, making it highly practical for multimodal datasets such as retinal fundus images and angiographic scans.

LS-CF was rigorously tested against multiple baselines, including thresholded Frangi, naive connectivity filters, and morphological closing, as well as a range of supervised and unsupervised literature approaches. On the OSIRIX angiographic dataset, it achieved the highest accuracy among all methods, outperforming every previously published work. On the IOSTAR dataset, it beat 4 out of 5 state-of-the-art methods, and on CHASE-DB, it surpassed all existing unsupervised techniques. Results on DRIVE and STARE were also competitive, demonstrating the filter's broad applicability. The work was published in the Journal of Imaging (2022) and later submitted to arXiv in 2026, signaling ongoing relevance for automated retinal disease screening and vascular analysis.