Exploring the Impact of Preprocessing Techniques on Retinal Blood Vessel Segmentation Using a Study Group Learning Scheme

Abstract

The segmentation of retinal vessels in retinal images is vital for automated diagnosis of retinal diseases. This is a challenging task because it requires accurate manual labeling of the vessels by expert clinicians and the detection of tiny vessels is difficult due to limited samples, low contrast, and noise. In this study, we explore the use of pre-processing techniques such as contrast-limited adaptive histogram equalization (CLAHE), grad-cam analysis and min-max contrast stretching to improve the performance of a study-group learning (SGL) segmentation model. We evaluate the impact of these pre-processing techniques on the accuracy, sensitivity, specificity, AUC, IoU, and Dice scores using four publicly available datasets, DRIVE, CHASE, HRF and IOSTAR. Our findings indicate that the utilization of the Min-Max technique resulted in a notable enhancement in the accuracy of both the DRIVE and CHASE datasets, with an approximate increase of 3% and 2% respectively. Conversely, the impact of the CLAHE method was discernible solely in the DRIVE dataset, demonstrating an improvement in accuracy of 1%. In addition, our results demonstrated superior accuracy performance for both the DRIVE and CHASE datasets compared to the findings of the reviewed studies. The GitHub repo for this project is available at Link.