Abstract
Purpose :
The purpose of this study was to develop a machine learning classification model to automatically classify glaucoma patients from normal subjects, with optical coherence tomography (OCT) and color fundus images.
Methods :
This study included 124 open-angle glaucoma and 180 normal eyes, and all those eyes were captured with spectral-domain OCT (3D OCT-2000, Topcon), in protocol of macular and disc 3 dimensional (3D) scans, as well as color fundus images.
With the exsting OCT segmentation sofware, we created retinal nerve fiber layer (RNFL) thickness and deviation map from disc 3D scans, and ganglion cell complex (GCC) thicknees and deviation map from macular 3D scans. With the created images from randomly selected subjects for training (n=204), we trained 5 separate machine learning classification model, and validated them with remaining data (n=100); 1) random forest (RF) for the color fundus images (green channel), 2) convolutional neural network (CNN) for disc RNFL thickness maps, 3) CNN for macular GCC thickness maps, 4) support vector machine (SVM) for disc RNFL deviation maps, 5) SVM for macular GCC deviation maps. Features for model 1) were extracted from a CNN model pre-trained on ImageNet, while for 4) and 5), features were created by non-negative matrix factorization method. Finally, another SVM model was trained to combine the five separate models to get a better classification accuracy, with the confidence from each model as the input.
Results :
The validated accuracy for each machine learning model was as below; 1) 88.0%, 2) 90.0%, 3) 91.0%, 4) 92.0%, 5) 92.0%, while for combining the results from five separate model, the accuracy was improved to 96.0%.
Conclusions :
This novel machine learning model was able to classify normal and glaucoma based on OCT and color fundus images with high and stable accuracy, and might be used in assisting glaucoma diagnosis.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.