Robustní segmentace obrazu pro automatickou detekci retinálního cévního systému z variabilních klinických obrazových dat

Abstract

Automatic segmentation of retinal blood vessels is essential for diagnosing numerous ocular diseases, yet the diversity of camera systems and variable image quality greatly undermine its reliability. This thesis therefore set out to optimise and evaluate a vessel‑segmentation algorithm. A gold‑standard database of 150 neonatal images captured with RetCam 3, Phoenix Icon and RetCam Envision was created, each image containing manual vessel annotations. The algorithm merges classical image‑processing steps (green‑channel extraction, LAB luminance adjustment, matched filtering, top‑hat transform) with Frangi and Jerman vessel filters. Forty parameter settings were tested on nine representative images, and performance was assessed with seven metrics, the key one being the Jaccard index. In evaluation, RetCam 3 yielded the highest average scores (accuracy 0.9829; Jaccard 0.4596; F‑score 0.6283), while Phoenix Icon and RetCam Envision showed comparable performance with median Jaccard values of about 0.28. Finally, a graphical software interface was implemented to enable straightforward use and adjustment of the segmentation algorithm.