Performance Evaluation of Different Object Detection Models for the Segmentation of Optical Cups and Discs

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Glaucoma diagnosis relies on optic disc and cup measurements, often using time-consuming deep learning segmentation.
• Object detection models offer a precise alternative for feature extraction from retinal fundus images.

Purpose of the Study:

• To compare the performance of various object detection models for automated segmentation of optic discs and cups in fundus images.
• To evaluate the effectiveness of models like Mask R-CNN, MS R-CNN, CARAFE, Cascade Mask R-CNN, GCNet, SOLO, and Point_Rend.

Main Methods:

• Models were evaluated on the Retinal Fundus Images for Glaucoma Analysis (REFUGE) and G1020 datasets.
• Performance metrics included Average Precision (AP), F1-score, Intersection over Union (IoU), and Area Under the Curve of Precision-Recall (AUCPR).

Main Results:

• Several models achieved perfect AP (1.000) on the REFUGE dataset with an IoU threshold of 0.50.
• Point_Rend achieved the highest AP (0.956) on the G1020 dataset, while SOLO performed lowest (0.906).
• Excellent performance was observed with as few as 100 images, with improvements from data augmentation and multi-scale handling.

Conclusions:

• Object detection models demonstrate high precision and recall for optic disc and cup segmentation, proving efficient and effective for glaucoma analysis.
• The models show promising cross-dataset transferability and effectiveness even with limited training data.