Purchase this article with an account.
Dominika Sulot, David Alonso-Caneiro, Patrycja Krzyzanowska-Berkowska, D. Robert Iskander; Differentiating glaucoma patients from healthy controls and glaucoma suspects using deep learning. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1648.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To design a deep learning model able to classify images of optic nerve head (ONH) from scanning laser ophthalmoscopy (SLO) belonging to three groups: healthy control subjects, glaucoma suspects and primary open-angle glaucoma patients.
A total of 266 images from an in-built OCT SLO (83 glaucoma suspects, 86 control subjects and 97 glaucoma patients) were used. The data set was divided into two groups: train and test set in a ratio of three to one. Due to the small amount of data, a custom convolutional neural network (CNN) architecture was designed. The results were obtained from a majority voting of five separately trained models on shifted sets of data received by moving the train and validation subsets inside the original train set (5-fold cross validation). Retinal nerve fiber layer (RNFL) thickness was measured in six ONH sectors. Additionally, gradient-based class activation maps (grad-CAMs) were calculated for every image and group statistics were evaluated in six different sectors of the ONH.
The proposed model achieved an overall accuracy of 86.6%. For individual groups, the model had the highest accuracy, sensitivity and specificity in differentiating glaucoma patients from other groups with score of 0.94, 0.93 and 0.95, respectively. For control subjects the model yielded: 0.93, 0.95 and 0.92 and for glaucoma suspects: 0.87, 0.70 and 0.94. Interestingly, correlation was found between grad-CAMs for glaucoma group and the differences in RNFL thickness between control and glaucoma groups (R=0.873, p=0.023).
The proposed deep learning method was found to be useful for differentiating glaucoma patients from control subjects and glaucoma suspects. Through data augmentation and model ensemble, the proposed method distinguishes three groups with high accuracy, despite a small data set. Correlation between grad-CAMs and RNFL thickness suggest that in future research more attention should be given to the interpretation of those maps and their relationships to other clinical parameters. This study also highlights the potential of non-structural SLO images in the detection of glaucoma.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only