Learning From AI-Generated Annotations for Medical Image Segmentation

Learning from AI-generated annotations is well-recognized as a key advance of deep learning techniques in medical image segmentation. Towards this direction, in this paper, we investigate two questions: (1) how to accurately measure loss value on AI-generated annotations that often contain errors and (2) how to effectively update model's parameters when the loss value is no longer a correct supervision for medical image segmentation. The main results are that (1) 'error-tolerant' loss functions exist and (2) 'cross-training', updating the model using data with a small loss of its 'twin' model, can tolerate the loss function to some extent. Per the main results, we yet derived a robust training algorithm, called confidence regularized co-teaching, that helps deep models to combat annotation errors in medical image segmentation. This algorithm simultaneously trains two 'twin' segmentation models and updates model's parameters by cross-training with disagreement confident data that are predicted differently by the two models, thereby being able to learning from data with annotation errors. The empirical evidence from a publicly available dataset shows that this new algorithm works better on combating annotation errors than existing methods for medical image segmentation, opening the opportunity to use AI-generated annotations to train segmentation model for medical image segmentation.