Code-Free Deep Learning Glaucoma Detection on Color Fundus Images.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Code-free deep learning (CFDL) empowers clinicians without coding expertise to develop AI models.
• Glaucoma detection from fundus images is crucial for early intervention and vision preservation.
• Evaluating CFDL performance against expert-designed models is essential for clinical adoption.

Purpose of the Study:

• To assess the performance of CFDL in detecting glaucoma from fundus images.
• To compare CFDL model performance with established expert-designed deep learning models.
• To validate the CFDL model's efficacy on external datasets.

Main Methods:

• Ophthalmology trainees developed a CFDL binary classification model using 101,442 labeled fundus images from the Rotterdam EyePACS AIROGS dataset.
• The CFDL model was trained to differentiate glaucoma from normal optic nerves.
• External validation was performed using the REFUGE and GAMMA datasets at varying confidence thresholds.

Main Results:

• The CFDL model achieved high performance metrics, including an Area Under the Precision-Recall Curve (AuPRC) of 0.988, 95% sensitivity at 95% specificity (SE@95SP), and 91% accuracy.
• Performance was comparable to bespoke deep learning models, with SE@95SP of 95% versus 85% for top bespoke models.
• External validation on REFUGE and GAMMA datasets demonstrated robust performance, with SE@95SP ranging from 83% to 98% and Area Under the Receiver Operating Curve (AUC) up to 0.994.

Conclusions:

• CFDL models show significant potential for glaucoma screening using fundus images, offering a viable alternative to traditional methods.
• This study provides a compelling proof of concept for CFDL in medical image analysis.
• CFDL empowers clinicians to innovate and develop tailored AI solutions for widespread glaucoma screening.