Robust optic disc and cup segmentation with deep learning for glaucoma detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Glaucoma is a leading cause of irreversible blindness globally.
• Early detection and treatment are vital for vision preservation.
• Accurate segmentation of optic disc and cup is essential for glaucoma diagnosis.

Purpose of the Study:

• To develop a robust optic disc and cup segmentation method using a modified U-Net architecture.
• To evaluate the model's performance on multiple datasets and compare it with expert performance and state-of-the-art methods.

Main Methods:

• A modified U-Net architecture combining a pre-trained ResNet-34 encoder with U-Net decoder was developed.
• The model was trained on the RIGA dataset and tested on RIGA, DRISHTI-GS, and RIM-ONE datasets.
• Performance was evaluated using dice similarity coefficients for disc and cup segmentation.

Main Results:

• The model achieved high dice values: 97.31% for disc and 87.61% for cup on the RIGA dataset.
• On unseen datasets (DRISHTI-GS, RIM-ONE), the model demonstrated state-of-the-art performance after fine-tuning.
• The combined ResNet-U-Net approach enabled faster training and robust performance, avoiding overfitting.

Conclusions:

• The proposed modified U-Net model offers a robust and efficient solution for optic disc and cup segmentation.
• This method shows potential for improving early glaucoma detection and management.
• The approach achieves state-of-the-art results, comparable to expert performance and existing methods.