Efficacy for Differentiating Nonglaucomatous Versus Glaucomatous Optic Neuropathy Using Deep Learning Systems.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Differentiating glaucomatous optic neuropathy (GON) from nonglaucomatous optic neuropathy with disc pallor (NGON) is crucial for appropriate patient management.
• Color fundus photography is a common imaging modality in ophthalmology.

Purpose of the Study:

• To evaluate the performance of deep learning algorithms in distinguishing between NGON and GON using color fundus photographs.
• To develop and assess an Artificial Intelligence Classification algorithm for optic neuropathy diagnosis.

Main Methods:

• A convolutional neural network (CNN) model, specifically ResNet-50 architecture, was employed.
• The study analyzed 3815 fundus images, including normal optic discs, NGON, and GON, with expert neuro-ophthalmologist corroboration.
• Image preprocessing involved size and color enhancement before AI model input.

Main Results:

• The ResNet-50 model achieved a sensitivity of 93.4% and specificity of 81.8% in detecting GON among NGON images.
• The area under the precision-recall curve (average precision) for differentiating NGON versus GON was 0.874.
• Identified false positive cases were associated with peripapillary atrophy and tilted optic discs.

Conclusions:

• AI-based deep learning models demonstrate excellent performance in differentiating NGON and GON from color fundus photographs.
• These findings suggest the potential clinical utility of AI in diagnosing optic disc diseases.
• Further research is warranted to validate these deep learning algorithms for widespread clinical application.