A New Approach for Fundus Lesions Instance Segmentation Based on Mask R-CNN X101-FPN Pre-Trained Architecture

Diabetic retinopathy is one of the main causes vision loss, and it can be identified through ophthalmological examinations that aim to locate presence retinal lesions such as Microaneurysms, Hemorrhages, Soft Exudates, Hard Exudates. The development computerized approaches perform instance segmentation these help in early diagnosis disease. However, instances artifacts retina a complex task due factors object size morphological characteristics. This article proposes new approach based on Mask Regions with Convolutional Neural Network features (Mask R-CNN) architecture associated diabetic retinopathy. proposed was trained, adjusted, tested using different public datasets retinopathy, which were implemented Detectron2 libraries OpenCV. best result obtained by Dataset for Retinopathy (DDR) Tilling Adam optimizer, reaching mean Average Precision ( $mAP$ ) 0.2903 detection fundus limit Intersection Over Union notation="LaTeX">$IoU$ 0.5 validation stage an 0.1670 test step. results experiments demonstrate presented promising microaneurysms increase precision, our case reaches approximately 16%.