Abstract
Computer vision technologies are increasingly being utilized to enhance the diagnostic processes of ophthalmic diseases. This paper presents a comprehensive software pipeline that incorporates image preprocessing, automated annotation generation, and object detection model training using YOLO. The primary goal is to improve the accuracy of pathological region localization by introducing a contrast enhancement stage based on CLAHE (Contrast Limited Adaptive Histogram Equalization) and analyzing its effect on model performance. Initially, the brightness levels of images from a retinal dataset are assessed. Low-brightness images are enhanced using CLAHE in the LAB color space to increase contrast without distorting color information. Binary masks and contour extraction techniques are then used to generate YOLO-format annotations. The YOLOv8 model is fine-tuned on a custom dataset using a version pre-trained on the REFUGE dataset. The performance evaluation includes prediction visualization, PRcurve plotting, confusion matrix analysis, and the IoU (Intersection over Union) metric. Experimental Results show that CLAHE preprocessing improves average IoU from 0.93 to 0.94, indicating enhanced localization accuracy. The proposed pipeline demonstrates the effectiveness of contrast enhancement methods in ophthalmic object detection tasks and can be extended to other types of medical images and tasks, such as segmentation or classification.