Diabetic Retinopathy Classification with pre-trained Image Enhancement Model

Diabetic Retinopathy is one of the main causes of blindness. The degree of retinopathy can be detected in images of the retinal fundus. Various machines and deep learning techniques are developed for automatic detection. However, a huge amount of training images is required to achieve a high-performance model, which does not exist in some domains. We proposed a hybrid training approach by including a trained knowledge base technique in traditional deep learning model training. The knowledge base model is created by an artificial expert, a simple deep learning model. The feature of interest is identified by a pre-trained model, and then the deep convolutional neural network is applied for image classification. Consequently, our approach requires a small number of training images and provides a model with higher performance compared with the baseline model.

​​​​​​​Keyword: Convolutional Neural Network, Deep Learning, Diabetic Retinopathy, Small Number of Training Set