Compact Convolutional Transformer For Classification Of Retinal Diseases From Optical Coherence Tomography Images

Abstract

In this study, we focus on how to use advanced deep learning models to identify retinal disorders from OCT images automatically. To increase the accuracy and speed of clinical diagnostics, the study introduces the Compact Convolutional Transformer (CCT) which is a transformer-based model that is both lightweight and performs well. A large OCT dataset made up of more than 130,000 images is used in the study to detect CNV, DME, Drusen and Normal conditions. A key part of our method is a data preparation process that uses image rotation, flipping and zooming to ensure classes are balanced. A 32×32 pixel image size was used for classification to lessen the computing burden without affecting how accurately the diagnosis was made. After extensive testing, the suggested model was matched against a standard Vision Transformer (ViT) model and a group of eight widely used transfer learning models, including DenseNet121, ResNet50 and EfficientNetB1. The model’s effectiveness was tested by performing evaluation with several measures, including precision, recall, F1 score and training time. The study showed that the CCT model is more effective and resilient than the baseline and transfer learning models in medical picture classification tasks. This study highlights how transformer-based models, in particular CCT, can be included into AI-assisted retinal diagnostic systems in real time