Deep learning system identifies arteritic, nonarteritic anterior ischemic optic neuropathy
Click Here to Manage Email Alerts
Key takeaways:
- The deep learning system demonstrated an AUC of 0.97, a sensitivity of 91.1%, a specificity of 93.4% and an accuracy of 92.6%.
- The system misclassified only 7.3% of images.
A deep learning system achieved an accuracy of more than 90% in distinguishing arteritic from nonarteritic anterior ischemic optic neuropathy using only color fundus images, according to a study published in JAMA Ophthalmology.
“Distinguishing arteritic anterior ischemic optic neuropathy (AAION) from nonarteritic anterior ischemic optic neuropathy (NAION) can be clinically challenging, given their similar clinical presentations (ie, sudden monocular visual loss associated with optic disc swelling),” Ayse Gungor, MSc, a PhD candidate at Sorbonne University and Rothschild Foundation Hospital in Paris, and colleagues wrote. “The rapid distinction of AAION from NAION is essential in clinical practice, because AAION demands urgent intervention with high-dose corticosteroids to prevent blindness of the fellow eye.”
Seeking to develop, train and test a dedicated deep learning system that could distinguish acute AAION from NAION using color fundus images, Gungor and colleagues conducted a study of 961 eyes from 802 patients, of whom 136 (16.9%) were diagnosed with AAION and 666 (83%) were diagnosed with NAION. Images from 21 neuro-ophthalmology centers in 16 countries were used for training and internal validation, while external testing was conducted at five neuro-ophthalmology centers in the U.S. and Europe.
According to results, the deep learning system demonstrated an area under the curve of 0.97, a sensitivity of 91.1%, a specificity of 93.4% and an accuracy of 92.6%. Of 136 images included in the external testing dataset, the system misclassified 10 images (7.3%) — four as NAION and six as AAION. The deep learning system prioritized the inferior part of the optic nerve head as a region of interest when distinguishing NAION from AAION.
The system also outperformed two independent neuro-ophthalmologists when making the distinction between AAION and NAION, with the experts achieving accuracies of 74.3% and 81.6% when classifying the external testing dataset.
“The [deep learning system] used in this study could discriminate between eyes with acute AAION and NAION on color fundus images without any additional clinical or biomarker information,” Gungor and colleagues wrote. “Such a system, if available in the clinical practice setting, could potentially improve the triaging process in older patients presenting with acute anterior ischemic optic neuropathy, in whom accurately discriminating between an arteritic vs. a nonarteritic etiology likely can have important therapeutic consequences.”
Collapse
Click Here to Manage Email Alerts