Multi-step framework for glaucoma diagnosis in retinal fundus images using deep learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of blindness globally.
• Deep learning models are commonly used for glaucoma screening from retinal fundus images.
• Blood vessels in the optic disc and pathology outside the optic disc can interfere with accurate deep learning-based glaucoma diagnosis.

Purpose of the Study:

• To propose a novel multi-step framework, MSGC-CNN, for enhanced glaucoma diagnosis.
• To improve diagnostic efficiency by integrating original fundus images with vessel-removed optic disc images.
• To address challenges in glaucoma fundus image analysis, including limited data, high resolution, and rich feature information.

Main Methods:

• A multi-step framework (MSGC-CNN) was developed, combining glaucoma pathological knowledge with deep learning.
• Features from original fundus images and U-Net-processed, vessel-removed optic disc regions were fused.
• A novel feature extraction network (RA-ResNet) was designed and combined with transfer learning to handle specific image characteristics.

Main Results:

• Binary classification experiments were conducted on three public datasets: Drishti-GS, RIM-ONE-R3, and ACRIMA.
• The MSGC-CNN framework achieved high accuracy rates of 92.01% on Drishti-GS, 93.75% on RIM-ONE-R3, and 97.87% on ACRIMA.
• The proposed method demonstrated significant improvements compared to previous results.

Conclusions:

• The MSGC-CNN framework effectively enhances glaucoma diagnosis by integrating diverse image features.
• The novel approach successfully overcomes limitations of traditional deep learning methods in glaucoma screening.
• The results indicate a promising advancement in automated glaucoma detection using medical imaging and AI.