基于多尺度区域可靠性感知的半监督肺肿瘤分割

Semi-supervised lung tumor segmentation based on multi-scale consistency and regional reliability perception

提出一种基于多尺度下一致性和区域可靠性感知的半监督学习方法,在使用少量标注数据的情况下结合未标记的数据来实现高性能的肺部肿瘤分割任务。首先,提出一种多尺度一致性均值教师框架,构建多尺度一致性损失并约束教师学生网络中多个尺度上的输出保持一致,使模型学习到更丰富的一致性知识。此外,提出一种区域可靠性感知方法使一致性学习之间的知识交换更加有效,使模型从无标注的数据中学习到更有效且可靠的知识。本文方法在医学图像分割十项全能比赛肺肿瘤数据集上进行充分的评估,与当前先进的半监督学习方法比较有更优越的性能,验证本文方法的有效性。

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.