Semi-supervised lung tumor segmentation based on multi-scale consistency and regional reliability perception
A semi-supervised learning method based on multi-scale consistency and regional reliability perception is proposed to combine unlabeled data with a small amount of labeled data to achieve high-performance lung tumor segmentation tasks.A multi-scale consistency mean teacher framework is used to construct a multi-scale consistency loss and constrain the outputs in the mean teacher network to be consistent across multiple scales,so that the model learns richer consistency knowledge.In addition,a regional reliability perception scheme is adopted to make the knowledge exchange between consistency learning more efficient,enabling the model to learn more valid and reliable knowledge from unlabeled data.The evaluation on the lung tumor dataset in the Medical Segmentation Decathlon shows superior performance of the proposed method over current state-of-the-art semi-supervised learning methods,validating its effectiveness.
NETL
NSTL
万方数据
