Small Scale Feature Propagation Using Deep Residual Learning for Diabetic Retinopathy Classification

Automated diagnosis of diabetic retinopathy from fundus images involves detecting both small- and large-scale lesions, which makes this a difficult task for deep learning applications. In this paper we investigate the effects of small scale feature propagation for improving diabetic retinopathy classification. To accomplish this, we have utilized a publicly available dataset with 88,702 images, which contains unbalanced number of examples for different classes. A linear equation for class-specific gradient weighting has been proposed and has found to be beneficial. Three different residual architectures with residual and skip connections have been tested and their efficacy for this task is examined. The residual connections have been verified to improve the results for detecting small scale features for deep architectures. Skip connections within the current experimental setting have been found to be detrimental for the overall performance, potential solutions and their resulting effects have been discussed.