Research on Retinal OCT Image Classification Based on Deep Learning

Abstract

Presently, there persists an issue in convolutional neural network (CNN)-based classification methods for retinal optical coherence tomography (OCT) images, particularly in discerning small-scale lesion areas. This dilemma results in neural network models exhibiting diminished accuracy when determining the activity of age-related macular degeneration (AMD) regarding its dry and wet stages, as well as choroidal neovascularization (CNV). Accurate identification of lesion types is paramount for ophthalmologists in devising treatment strategies. To address this challenge, a transfer learning-based EfficientNet retinal OCT image classification algorithm is proposed. Initially, retinal OCT images undergo data augmentation and preprocessing procedures. Subsequently, the pre-trained EfficientNet-B3 model is trained via transfer learning, followed by fine-tuning training using partial oversampling and class weighting techniques. The ultimate classification accuracy reaches 99.2%, signifying the model's commendable classification recognition accuracy. This underscores the clinical guidance significance of this research endeavor.