Optical coherence tomography retinal classification using deep neural network

By detecting curable retinal diseases in their early stages, optical coherence tomography (OCT) has helped people avoid or significantly lessen the severity of irreversible vision deterioration or completely losing it. Using classical convolution neural networks (CNN), it is possible to get very accurate classification of OCT images. But conventional convolutional CNN models have taken a lot of heat for their pooling layers, which supposedly bury positional interactions. Since the typical VGG16 model uses a smaller number of pooling layers, it was chosen for this work. After making some adjustments to this basic VGG16 model, we will apply a transfer learning technique for the training-phase of the resultant model. The OCT dataset has been divided into 70% for training, 20% for the testing, and 10% for the validation phase, where this validation part will never be being seen by the system during training and testing phases. Our method outperformed other methods outlined in the available literature for OCT image classification, achieving an accuracy of 97%.