CMRI Segmentation Domain Generalization by Random Style Transfer
Haixin Song, Lei Li, Zhiming Luo, Sheng Lian, Shaozi Li · 2022
When training data and test data have different distribution, performance of neural networks may decrease significantly, and model is difficult to generalize knowledge. Domain generalization (DG) attempts to learn general and transferable knowledge from a variety of source domain data to enable the model to generalize to unseen domains. In this paper, considering both DG and few-shot scenario, we propose a random stylization block to augment existing styles through nonlinear combination of data, called style escape, so as to obtain more unseen styles for the model to learn consistency. We propose a heuristic gradient weighting strategy to optimize the training process of meta-learning model. The experimental results on M&Ms dataset show that segmentation performance of our model is significantly improved compared with baselines, and the model training cost is noticeably reduced, which fully verifies the effectiveness of our method.