Accuracy of machine learning for differentiation between optic neuropathies and pseudopapilledema

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Differentiating optic neuropathies and pseudopapilledema (PPE) from normal optic discs is clinically significant.
• Accurate diagnosis is crucial for appropriate patient management and preventing vision loss.

Purpose of the Study:

• To evaluate the diagnostic accuracy of machine learning (ML) algorithms in distinguishing between optic neuropathies, PPE, and normal optic discs.
• To compare the performance of various ML classifiers using fundus photography.

Main Methods:

• A dataset of 295 optic neuropathy images, 295 PPE images, and 779 control images was utilized.
• Four ML classifiers were compared: a custom model, GoogleNet Inception v3, VGG, and ResNet.
• Performance was assessed using accuracy and area under the receiver operating characteristic curve (AUROC).

Main Results:

• ML classifier accuracy ranged from 95.89% to 98.63%.
• ResNet achieved the highest accuracy (98.63%), followed by VGG (96.80%), Inception v3 (96.45%), and the custom model (95.89%).
• ResNet and VGG demonstrated excellent diagnostic capability with an AUROC of 0.999.

Conclusions:

• Machine learning, combined with fundus photography, offers an effective method for differentiating PPE from optic disc elevation due to optic neuropathies.
• Deep learning models show significant promise in improving the diagnostic accuracy of optic disc abnormalities.