Machine learning of retinal pathology in optical coherence tomography images

Pushkar Aggarwal


Background: Acute macular degeneration (AMD), central serous retinopathy (CSR), diabetic retinopathy (DR) and macular hole (MH) are common vision impairing pathologies in the field of ophthalmology. Machine learning with deep convolutional neural networks can be used to analyze ophthalmological diseases using fundus and optical coherence tomography (OCT) images, but with limited accuracy. In order to improve the sensitivity and specificity of these models, the objective of this study was to examine the effect of data augmentation on the performance of the neural network.
Methods: OCT Images for above pathologies and normal eye were acquired from the Optical Coherence Tomography Image Database. Keras, a neural network framework, was used to retrain Visual Geometry Group 16 (VGG16), a deep neural network, using these images. Retraining was performed with and without data augmentation on two separate models. Data augmentation techniques included rotation, shear, horizontal flip and Gaussian noise.
Results: Average Matthews correlation coefficient (MCC) increased from 0.83 in the model without data augmentation to 0.93 in the model with data augmentation. Average statistical measures- sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), MCC and F1 score increased with data augmentation. The average area under the curve (AUC) increased from 0.91 to 0.97 with data augmentation addition.
Conclusions: Data augmentation techniques can be used in machine learning to appreciably increase the accuracy of a deep convolutional neural network. In future applications, the model created in this analysis can be retrained with a higher quantity and better quality of images and provided to physicians as an aid when examining OCT images.