Purchase this article with an account.
vahid mohammadzadeh, Alex Broering, Arvind Vepa, Zehao Pan, Alon Oyler-yaniv, Diana Salazar, Golnoush Sadat Mahmoudi Nezhad, Jack Martinyan, Esteban Morales, Joseph Caprioli, Fabien Scalzo, Kouros Nouri-Mahdavi; Detection of Glaucoma Progression on Optic Disc Photographs (DP) with Deep Learning. Invest. Ophthalmol. Vis. Sci. 2020;61(7):873.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop supervised deep learning (DL) techniques for detection of change in a group eyes with longitudinal stereoscopic DPs and to validate their performance based on clinical review of DPs by experienced clinicians (‘ground truth’).
DPs from 1647 eyes (916 patients) with suspected or established glaucoma at baseline, 2 DPs and 2 years of follow-up were included. Baseline and final DPs were evaluated for progression by 2 clinicians and the results were adjudicated by two glaucoma specialists. A DL transfer learning model using a pre-trained ImageNet dataset was used to distinguish right and left eyes. To improve the generalizability of the model, data augmentation such as blur, rotation, cropping, brightness augmentation, horizontal and vertical flips, as well as a black border (to simulate digitized slides) was used to increase the input 10-fold. A U-Net style model for segmentation of the optic disc cup was trained using 137 labeled images. The base encoder was a SE-ResNeXt50 with 32×4D template, pre-trained on ImageNet. Similar augmentations to classification were used for segmentation. For the classification model, a custom Convolutional Neural Network (CNN) with a sigmoid classifier was used. The model has two identical CNN branches through which the base and final image for each eye are entered before concatenation of results and final classification. Additionally, masks were used to weight the convolutions and help the CNN focus on the optic disc and peripapillary region, where progressive changes in glaucoma can be detected. The area under ROC curves was the main metric used to evaluate the performance of the model given the imbalance between the two classes.
Median (IQR) follow-up time was 7.8 (2.4-13.4) years. The average (±SD) visual field mean deviation was –4.6 (±6.1) at baseline. 22% of DP series were considered to be progressing on clinical review. The optimized DL approach identified 18% of eyes as progressing. The agreement with the clinical review (accuracy) was 93 %. The area under the ROC curve was 0.84.
Our optimized DL algorithm was able to detect glaucoma progression based on longitudinal DPs with very good performance. Implementation of this DL model, after external validation, could significantly enhance the challenging task of detection of glaucoma progression and make this task more efficient for clinicians.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only