Computer-aided recognition of myopic tilted optic disc using deep learning algorithms in fundus photography

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Myopic optic disc tilt significantly affects ocular parameters.
• Current ophthalmologic measurements are prone to inter-observer variability and are time-consuming.
• Automated detection of myopic optic disc tilt is needed to improve diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate deep learning models for automatic recognition of myopic optic disc tilt in fundus photography.
• To compare the performance of deep learning models against human expert evaluation.
• To assess the clinical utility of an automated system for identifying optic disc tilt.

Main Methods:

• Utilized 937 fundus photographs from patients with normal or myopic tilted discs.
• Developed a deep learning system using GoogleNet Inception-v3 architecture.
• Employed image resizing techniques and data augmentation for model training and evaluation.
• Analyzed model performance using metrics like AUC, accuracy, sensitivity, and specificity.
• Visualized model decision-making using Grad-CAM++.

Main Results:

• The deep learning models achieved an AUC of 0.978 ± 0.008, accuracy of 0.960 ± 0.010, sensitivity of 0.937 ± 0.023, and specificity of 0.963 ± 0.015.
• Simple image resizing and augmentation yielded the best model performance.
• Activation map visualization confirmed the model's ability to identify key ocular structures like optic discs and maculae.
• The system demonstrated excellent agreement with clinical criteria for optic disc tilt.

Conclusions:

• An automated deep learning system for detecting optic disc tilt has been successfully developed.
• The model exhibits high accuracy and reliability, comparable to expert clinicians.
• This technology holds promise for future applications in ophthalmology, aiding in the adjustment and identification of optic disc tilt effects on measurements.