Transfer Learning-Based Classification of Optical Coherence Tomography Retinal Images

Currently, optical coherence tomography is one of the most sensitive methods for detecting diabetic retinopathy. However, the artificial detection of diabetic retinopathy is time consuming and prone to subjective errors. Accordingly, this paper proposed an improved deep learning network based on transfer learning for automatic classification of retinal images. First, the image was preprocessed via adaptive threshold combined with the Gaussian filter algorithm. Then, on the basis of the pretraining model, the problem of sample difference was solved through fine-tuning, and the traditional fully connected layer was replaced by the global average pooling method for extracting deep features and reducing overfitting. The network was validated based on the experimental data, with the accuracy of the retinal image classification being 97.3%. Results reveal that the proposed network is effective for the automatic classification of retinal macular lesions.