OCTModNet: a deep learning-based framework for optical coherence tomography image multi-class classification

Abstract

Sight is one of the most significant senses in human beings. Losing sight can change a human’s life dramatically. Early diagnosis and detection of retinal diseases will prevent sight loss. Optical coherence tomography (OCT) imaging is helpful in ocular imaging with its high resolution and non invasive imaging for diagnosing retinal diseases. This study proposes OCTModNet, a novel multi-class hybrid classification model that is capable of classifying seven different retinal diseases using a combination of publicly available datasets, Kaggle OCT-C8, and OCTDL datasets. For effective extraction of the significant features, the collected data is resized and normalized. Augmentation techniques are used to improve the training and learning process. For model development, we applied transfer learning and fine-tuned several state-of-the-art deep learning models, ResNet50, InceptionV3, DenseNet121, VGG16, and EfficientNetV2S. Two models that performed best, VGG16 and EfficientNetV2S, were selected as the model backbones with integration of a squeeze and excitation block. The results showed that the OCTModNet achieved an outstanding performance with 99% test accuracy, 99% precision, 99% recall, 99% F1-score, and an area under the curve (AUC) as 99% across the unseen data. These results point out the robustness and reliability of the proposed model to classify OCT images and have the potential to enhance clinical decision-making and assist ophthalmologists in the early detection of diseases.