Multi-scale Stepwise Training Strategy of Convolutional Neural Networks for Diabetic Retinopathy Severity Assessment

Diabetic retinopathy severity assessment is an important domain in which deep learning has benefited medical imaging analysis. In this regard, CNNs which perform well in ImageNet are incapable of extracting subtle lesion features from high-resolution retinal fundus images. So novel convolutional networks with higher input size were developed. But no prior work give deep investigation on the impact of image resolution in the context of DR severity assessment. In this paper, we first explore how the performance of diabetic retinopathy severity assessment task would change if higher-resolution input images were used. Next, we adopt the stepwise strategy of training convolutional networks with high input scales to avoid overfitting. Finally, rigorous analyses on the impact of image resolution are given, showing that as model expands with higher input image resolutions, the performance grows logarithmically while both time and space complexity increase exponentially. Our model obtains new state-of-the-art kappa score in the task of diabetic retinopathy severity assessment task on EyePACS dataset with convolutional networks whose input size is 896 x 896, and great progress in classification of mild diabetic retinopathy. There is great potential for generalizing this solution to other medical image analysis problems.