The Application Of Deep Learning In The Field Of Retinal Diseases

Purpose:To train a model to classify multimodal and multi-categorical retinal images,so as to diagnose retinal diseases.Method:We used a transfer learning approach to train a deep convolutional neural network(CNN)with a dataset of 14057 retinal images classified into 20 categories.Images of 5 different modalities come from two different devices including Zeiss FF450~pluslus and Optos 200Tx.Result:Our model achieved a high level of performance on validation set and test set.The sensitivity,specificity and AUC was 94.8%,100%,0.99,respectively,on validation set.The sensitivity,specificity and AUC was 90.5%,100%,0.99,respectively,on test set.On MESSIDOR data set,the sensitivity is 62.3%,specificity is 100%.On E-Ophtha data set,sensitivity is 70.8%,specificity is 100%.Conclusion:This study showed that variations in images did not affect model’s performance.On the contrary,images from different modalities might help to improve feature representation,reduce the demand of data and overfitting.Despite trained on a relatively small data set,our model achieved a high level of performance on validation set and test set.Possible contributions might be the use of transfer learning and feature sharing in different images.