Leveraging the Generalization Ability of Deep Convolutional Neural Networks for Improving Classifiers for Color Fundus Photographs

Deep learning demands a large amount of annotated data, and the annotation task is often crowdsourced for economic efficiency. When the annotation task is delegated to non-experts, the dataset may contain data with inaccurate labels. Noisy labels not only yield classification models with sub-optimal performance, but may also impede their optimization dynamics. In this work, we propose exploiting the pattern recognition capacity of deep convolutional neural networks to filter out supposedly mislabeled cases while training. We suggest a training method that references softmax outputs to judge the correctness of the given labels. This approach achieved outstanding performance compared to the existing methods in various noise settings on a large-scale dataset (Kaggle 2015 Diabetic Retinopathy). Furthermore, we demonstrate a method mining positive cases from a pool of unlabeled images by exploiting the generalization ability. With this method, we won first place on the offsite validation dataset in pathological myopia classification challenge (PALM), achieving the AUROC of 0.9993 in the final submission. Source codes are publicly available.