Diabetic Retinopathy Features Segmentation without Coding Experience with Computer Vision Models YOLOv8 and YOLOv9

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, characterized by early signs like microaneurysms (MAs), hemorrhages (HEMOs), and exudates (EXs).
• Computer vision models are increasingly utilized for automated detection and classification of these early DR indicators in retinal fundus images.

Purpose of the Study:

• To evaluate the performance of YOLOv8 and YOLOv9 architectures for segmenting DR-related lesions and the optic disc.
• To assess the feasibility of using these models for DR lesion detection without requiring extensive coding or programming expertise.

Main Methods:

• Utilized 100 DR fundus images from the MESSIDOR database, manually annotated for pixel segmentation.
• Applied data augmentation techniques including tiling, flipping, and rotating to enhance training sample diversity.
• Tested various YOLOv8 and YOLOv9 model variants for lesion detection and segmentation.

Main Results:

• Achieved acceptable mean average precision (mAP) in detecting DR lesions (MA, HEMO, EX) and the optic disc.
• Demonstrated the potential of YOLO architectures for automated analysis of DR fundus images.
• Compared performance against other neural network approaches in the literature.

Conclusions:

• YOLOv8 and YOLOv9 show promising results for DR lesion segmentation, but are not yet ready for clinical implementation.
• Accurate lesion detection is crucial for early and correct diagnosis of diabetic retinopathy.
• Future work should focus on improving MA segmentation, image pre-processing, and utilizing standardized datasets.