Multiclass Classification of Retinal Disorders Using Optical Coherence Tomography Images

The human eye is crucial for interacting with the world, yet it is vulnerable to retinal diseases that can progressively impair vision and potentially lead to blindness if untreated. Accurate and early detection is essential for effective treatment and prevention. This study aims to develop an automated system for detecting and classifying retinal disorders using Optical Coherence Tomography (OCT) images. Four models were utilized: 1) OCTNet (a custom convolutional neural network (CNN)), 2) a modified version of VGG-19, 3) Inception V3, and 4) ResNet101, implemented on both original and preprocessed data. The usage of Gaussian median wavelet filters dealt with data preprocessing. OCT-specific techniques such as contrast stretching, histogram equalization, and adaptive histogram equalization are used for image improvement. OCTNet and the modified VGG-19 showed more accuracy as compared to other models that were being used for comparison. The dataset, sourced from Kaggle, consisted of 84,495 retinal OCT images sorted into four classes: CNV, DME, DRUSEN, and NORMAL. All the used models performed way better on preprocessed data than original data. OCTNet got the highest accuracy at 99%, with precision, recall, and F1 scores also matched this performance. The modified VGG-19 model scored an accuracy of 97%, with high precision, recall, and F1 scores. The test result shows that OCTNet is handy and very effective for detecting and classifying retinal disorders and getting the highest accuracy. This model can be very helpful in diagnosing the disease at early stages, which can also help in timely and accurate treatment to prevent severe vision loss. Future research in this area, by focusing on expanding the dataset, includes a broader range of retinal diseases and more refined models to improve their accuracy and generalizability.