Development of preprocessing methods and revised EfficientNet for diabetic retinopathy detection

The evolution of deep learning (DL) has made artificial intelligence image recognition a mature technology. Recently, the use of DL to identify diabetic retinopathy (DR) has been recognized as a major challenge. Retinal abnormalities caused by DR can damage the retina and thus cause permanent damage or even blindness. Therefore, the detection of diabetes symptoms at an early stage can help to considerably reduce the risk of blindness. The differences in environments, equipment, and photographers have led to the inconsistent specifications of images and have thus affected the efficiency of the training model for classifying the DR level. If low-quality fundus images are removed through an image quality assessment (Eye-Quality Library, EyeQ), the number of images of the trained model would considerably reduce, in turn affecting the reliability of the training model. To solve this problem, this paper proposes a preprocessing method to strengthen the image features. The results obtained in this study revealed that the preprocessing method could increase the amount of data available for the training model. Thus, this study improved the EfficientNet model for the enhancement of the classification performance of the DR level. The results also showed that an increase in the model accuracy from 0.7727 to 0.7920 for the classification of the different stages of DR. In addition, the results revealed that the revised EfficientNet could obtain better average area under the ROC curve among the five classes (0.926) than MobileNet (0.54) and the original EfficientNet (0.922). Finally, this study implemented the proposed system by using an application programming interface (API) to enable the users to upload a fundus image to the API and obtain the DR results.