Solution to OCT Diagnosis Using Simple Baseline CNN Models and Hyperparameter Tuning

Aim: Computer-aided diagnosis of retinopathy is a research hotspot in the field of medical image classification. Optical coherence tomography (OCT) is widely applied in the diagnosis of ocular diseases. In this study, we aim to classify the images into four classes, namely CNV, DME, drusen and normal. Methodology: In this study, we present the solution to classify the OCT images using simple baseline 3, 5 and 7 layer deep convolutional neural networks (CNNs). It also explores the effect of hyperparameters such as dropout, image size, batch normalisation, epochs and their relationships with the accuracy, sensitivity and specificity of the models. Results: The novelty of this study is that it does not use any pre-trained models and yet achieves desired results just by hyperparameter tuning and some clever observations. The best results were yielded by 5 layer model having hyperparameters image size 64 × 64, 30 epochs with dropout and batch normalisation achieving an accuracy of 97.92%. The biggest risk of overfitting in deep learning where multiple layered models are trained and tested and approaches of diminishing the overfitting effect has been discussed in detail.