Feature Perturbations Extend the Space for Automatic Grading of Diabetic Retinopathy

Contrast Learning (CL) is one of the most successful paradigms in Self-Supervised Learning (SSL) and has gained a lot of attention in the field of Deep Learning because of its strong visual representations, which do not require extensive labeling of the dataset, thus avoiding expensive costs. Currently, CL only explores the image space, neglecting the transformation of the feature representation. In this work, we propose a perturbed CL method that performs feature perturbation on the features extracted from the underlying network during training, extending the perturbation space of CL to better adapt the model to different data variations and to improve the generalization ability and robustness. We demonstrate that the proposed method can effectively improve the accuracy of diabetic retinopathy grading when evaluating the transmission capacity on the publicly available dataset EyePACS.