Two-stage framework for optic disc localization and glaucoma classification in retinal fundus images using deep...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Computer-aided diagnosis (CAD) is increasingly vital in medicine, particularly ophthalmology, for large-scale disease screening.
• Accurate detection and classification of the optic disc in retinal fundus images are crucial for glaucoma diagnosis.

Purpose of the Study:

• To propose a novel two-stage deep learning framework for automated optic disc detection, localization, and glaucoma classification.
• To address the lack of bounding box annotations in public datasets by developing a semi-automatic ground truth generation method.

Main Methods:

• A two-stage approach utilizing Regions with Convolutional Neural Network (RCNN) for optic disc localization and Deep Convolutional Neural Network (DCNN) for glaucoma classification.
• Development of a rule-based semi-automatic method for generating bounding box ground truth annotations for training the RCNN model.

Main Results:

• The proposed optic disc localization method achieved state-of-the-art results on six public datasets, with 100% accuracy on four.
• Glaucoma classification on the ORIGA dataset yielded an Area Under the ROC Curve (AUC) of 0.874, a 2.7% relative improvement over existing methods.

Conclusions:

• Deep learning models, once trained, offer robust, accurate, and automated solutions for optic disc analysis, eliminating the need for heuristic algorithms.
• Performance evaluation for imbalanced datasets necessitates additional metrics beyond AUC to fully represent classifier performance.