Classification of optical coherence tomography images using deep machine-learning methods

Abstract

BACKGROUND: Optical coherence tomography is a modern high-tech, insightful approach to detecting pathologies of the retina and preretinal layers of the vitreous body. However, the description and interpretation of study findings require advanced qualifications and special training of ophthalmologists and are highly time-consuming for both the doctor and the patient. Moreover, mathematical models based on artificial neural networks now allow for the automation of many image processing tasks. Therefore, addressing the issues of automated classification of optical coherence tomography images using deep learning artificial neural network models is crucial. AIM: To develop architectures of mathematical (computer) models based on deep learning of convolutional neural networks for the classification of retinal optical coherence tomography images; to compare the results of computational experiments conducted using Python tools in Google Colaboratory with single-model and multimodel approaches, and evaluate classification accuracy; and to determine the optimal architecture of models based on artificial neural networks, as well as the values of the hyperparameters used. MATERIALS AND METHODS: The original dataset included 2,000 anonymized optical coherence tomography images of real patients, obtained directly from the device with a resolution of 1,920×969×24 BPP. The number of image classes was 12. To create the training and validation datasets, a subject area of 1,100×550×24 BPP was “cut out”. Various approaches were studied: the possibility of using pretrained convolutional neural networks with transfer learning, techniques for resizing and augmenting images, and various combinations of the hyperparameters of models based on artificial neural networks. When compiling a model, the following parameters were used: Adam optimizer, categorical_crossentropy loss function, and accuracy. All technological operations involving images and models based on artificial neural networks were performed using Python language tools in Google Colaboratory. RESULTS: Single-model and multimodel approaches to the classification of retinal optical coherence tomography images were developed. Computational experiments on the automated classification of such images obtained from a DRI OCT Triton tomograph using various architectures of models based on artificial neural networks showed an accuracy of 98–100% during training and validation, and 85% during an additional test, which is a satisfactory result. The optimal architecture of the model based on an artificial neural network, a six-layer convolutional network, was selected, and the values of its hyperparameters were determined. CONCLUSION: Deep training of convolutional neural network models with various architectures, as well as their validation and testing, resulted in satisfactory classification accuracy of retinal optical coherence tomography images. These findings can be used in decision support systems in ophthalmology.