Glaucoma Detection and Feature Visualization from OCT Images Using Deep Learning

PurposeIn this paper, we aimed to clinically interpret Temporal-Superior-Nasal-Inferior-Temporal (TSNIT) retinal optical coherence tomography (OCT) images in a convolutional neural network (CNN) model to differentiate between normal and glaucomatous optic neuropathy. MethodsThree modified pre-trained deep learning (DL) models: SqueezeNet, ResNet18, and VGG16, were fine-tuned for transfer learning to visualize CNN features and detect glaucoma using 780 segmented and 780 raw TSNIT OCT B-scans of 370 glaucomatous and 410 normal images. The performance of the DL models was further investigated with Grad-CAM activation function to visualize which regions of the images are considered for the prediction of the two classes. ResultsFor glaucoma detection, VGG16 performed better than SqueezeNet and ResNet18 models, with the highest AUC (0.988) on validation data and accuracy of 93% for test data. Moreover, identical classification results were obtained from raw and segmented images. For feature localization, three models accurately identify the distinct retinal regions of the TSNIT images for glaucoma and normal eyes. ConclusionThis evidence-based result demonstrates the remarkable effectiveness of using raw TSNIT OCT B-scan for automated glaucoma detection using DL techniques which mitigates the black box problem of artificial intelligence (AI) and increases the transparency and reliability of the DL model for clinical interpretation. Moreover, the results imply that the raw TSNIT OCT scan can be used to detect glaucoma without any prior segmentation or pre-processing, which may be an attractive feature in large-scale screening applications.