A High-Accuracy Segmentation Hybrid Method For Retinal Blood Vessel Detection In Fluorescein Angiography Images Of Real Diabetic Retinopathy Patients

Abstract

This study aims to develop an advanced automatic method to enhance the accuracy of blood vessel segmentation in Fundus Fluorescein angiography FFA images of patients with real Diabetic retinopathy (DR) during both early and late phases. A prospective randomized study of 280 eyes of 87 DR patients at various stages was included in this study. our approach involved using four different image enhancement techniques (Histogram equalization (HE), CLAHE, RMSHE, Proposed method (CLAHE+RMSHE)) and for each enhancement except HE, four thresholding methods including the proposed method (Modified Active Contour(MAC)+ IsoData), C-mean Fuzzy, Isodata, and MAC + otsu thresholding that are commonly used in eye fundus images across different datasets. The effectiveness of segmentation was evaluated using three metrics: Dice (early: 0.84±0.05, late:0.84±0.03), Jaccard (early:0.73±0.06, late:0.74±0.05), and BF score (early:0.98±0.02, late:0.97±0.02). We found that the RMSHE method was more effective in the late phase, while the combined CLAHE+RMSHE method worked better in the early phase. The best segmentation results for both early and late phases were achieved using a blood vessel segmentation approach based on the MAC + IsoData thresholding.