Myopic-Net: Deep Learning-Based Direct Identification of Myopia Onset and Progression.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Myopia onset and progression (MOP) monitoring is crucial for clinical applications.
• Detecting MOP relies on analyzing anatomical changes in fundus retinal images.
• Deep neural networks (DNNs) offer a promising approach for automated analysis.

Purpose of the Study:

• To evaluate the performance of DNNs in identifying and monitoring MOP.
• To develop and validate a DNN model for detecting MOP from retinal images.

Main Methods:

• A DNN named Myopic-Net was developed using 6344 fundus image pairs (3964 without MOP, 2380 with MOP).
• Model performance was assessed using accuracy, precision, recall, specificity, and F1-score on internal and external test sets.
• Deep network visualization was employed to understand prediction drivers.

Main Results:

• Myopic-Net achieved 87.3% accuracy on the internal test set, surpassing human ophthalmologists (66.1% and 73.5%).
• The model maintained 84.1% accuracy on an independent external test set.
• Key predictive factors identified were anatomical changes in the optic disc and surrounding areas.

Conclusions:

• Myopic-Net effectively identifies MOP from fundus images by analyzing optic disc changes.
• The model demonstrates high accuracy, reliability, and generalization capabilities.
• DNNs show significant potential for monitoring and diagnosing MOP using fundus image analysis.