Machine learning assisted decision support system for image analysis of OCT

Abstract: Optical Coherence Tomography (OCT) has been around for more than 30 years and is still being continuously improved. The department of ophthalmology is a part of Sahlgrenska Hospital that heavily uses OCT for helping people with the treatment of eye diseases. They are currently facing a problem where the time to go from an OCT scan to treatment is being increased due to having an overload of patient visits every day. Since it requires a trained expert to analyze each OCT scan, the increase of patients is too overwhelming for the few experts that the department has. It is believed that the next phase of this medical field will be through the adoption of machine learning technology. This thesis has been issued by Sahlgrenska University Hospital (SUH), and they want to address the problem that ophthalmology has by introducing the use of machine learning into their workflow. This thesis aims to determine the best suited CNN through training and testing of pre-trained models and to build a tool that a model can be integrated into for use in ophthalmology. Transfer learning was used to compare three different types of pre-trained models offered by Keras, namely VGG16, InceptionResNet50V2 and ResNet50V2. They were all trained on an open dataset containing 84495 OCT images categorized into four different classes. These include the three diseases Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), drusen and normal eyes. To further improve the accuracy of the models, oversampling, undersampling, and data augmentation were applied to the training set and then tested in different variations. A web application was built using Tensorflow.js and Node.js that the best-performed model later was integrated into. The VGG16 model performed the best with only oversampling applied out of the three. It yielded an average of 95% precision, 95% recall and got a 95% F1-score. The second was the Inception model with only oversampling applied that got an average of 93% precision, 93% recall and a 93% F1-score. Last came the ResNet model with an average of 93% precision, 92% recall and a 92% F1-score. The results suggest that oversampling is the overall best technique for this given dataset. The chosen data augmentation techniques only lead to models performing marginally worse in all cases. It also suggests that pre-trained models with more parameters, such as the VGG16 model, have more feature mappings and, therefore, achieve higher accuracy. On this basis, parameters and better mappings of features should be taken into account when using pre-trained models.