Meeting News

Deep learning algorithm, fundus photos diagnose glaucoma

Researchers developed a deep learning algorithm to automatically diagnose glaucoma using a fundus photograph, according to a study presented at the Association for Research in Vision and Ophthalmology meeting.

“Our study suggests that a leading edge, deep learning method of the residual network achieved an equivalent or superior diagnostic performance of glaucoma from a fundus photograph as compared to medical doctors – even with a reasonably large training dataset and in highly myopic eyes,” researcher Ryo Asaoka, MD, PhD, said in a press release from ARVO.

The training dataset was made up of 1,364 color fundus photographs with glaucomatous appearances and 1,768 color fundus photographs without glaucomatous appearances.

The testing dataset consisted of: 34 eyes of 34 non-highly myopic glaucoma patients (group G), 30 eyes of 30 highly myopic glaucoma patients (group mG), 28 eyes of 28 non-highly myopic normative subjects (group N) and 22 eyes of 22 highly myopic normative subjects (group mN).

Fundus photographs were obtained using a nonmydriatic fundus camera (nonmyd WX3D, Kowa Ltd., Tokyo).

Researchers developed a deep learning algorithm of the residual network (resnet) to automatically diagnose glaucoma via a fundus photo. They tested the diagnostic accuracy of the resnet with the testing dataset, using the area under the receiver operating characteristic curve (AROC).

The AROC in all data was 95.4%. The sensitivity at the specificity of 95% was 70.3%, according to researchers. The AROC was 97.2% between groups G and N and 94.1% between groups mG and mN.

This technology may help screen more patients at a faster rate and identify disease prior to significant vision loss, according to the release. – by Abigail Sutton

Reference:

Asaoka R, et al. Construction of a deep learning algorithm to automatically diagnose glaucoma using a fundus photograph. Presented at: Association for Research in Vision and Ophthalmology; Honolulu; April 29-May 3, 2018.

Disclosures: Asaoka reported no relevant financial disclosures. Please see the full study for all remaining authors’ financial disclosures.

Researchers developed a deep learning algorithm to automatically diagnose glaucoma using a fundus photograph, according to a study presented at the Association for Research in Vision and Ophthalmology meeting.

“Our study suggests that a leading edge, deep learning method of the residual network achieved an equivalent or superior diagnostic performance of glaucoma from a fundus photograph as compared to medical doctors – even with a reasonably large training dataset and in highly myopic eyes,” researcher Ryo Asaoka, MD, PhD, said in a press release from ARVO.

The training dataset was made up of 1,364 color fundus photographs with glaucomatous appearances and 1,768 color fundus photographs without glaucomatous appearances.

The testing dataset consisted of: 34 eyes of 34 non-highly myopic glaucoma patients (group G), 30 eyes of 30 highly myopic glaucoma patients (group mG), 28 eyes of 28 non-highly myopic normative subjects (group N) and 22 eyes of 22 highly myopic normative subjects (group mN).

Fundus photographs were obtained using a nonmydriatic fundus camera (nonmyd WX3D, Kowa Ltd., Tokyo).

Researchers developed a deep learning algorithm of the residual network (resnet) to automatically diagnose glaucoma via a fundus photo. They tested the diagnostic accuracy of the resnet with the testing dataset, using the area under the receiver operating characteristic curve (AROC).

The AROC in all data was 95.4%. The sensitivity at the specificity of 95% was 70.3%, according to researchers. The AROC was 97.2% between groups G and N and 94.1% between groups mG and mN.

This technology may help screen more patients at a faster rate and identify disease prior to significant vision loss, according to the release. – by Abigail Sutton

Reference:

Asaoka R, et al. Construction of a deep learning algorithm to automatically diagnose glaucoma using a fundus photograph. Presented at: Association for Research in Vision and Ophthalmology; Honolulu; April 29-May 3, 2018.

Disclosures: Asaoka reported no relevant financial disclosures. Please see the full study for all remaining authors’ financial disclosures.

    See more from ARVO