【摘要】 针对医学图像标注成本高昂、耗时而半监督医学图像分割无标签数据分割不够精确、图像边缘信息损失、参数更新延迟等问题，提出一种基于一致性正则化的半监督医学图像分割方法。首先，设计一种基于熵-方差双视角不确定性度量方法来衡量无标签数据预测的不确定度，从熵和方差的角度联合评估无标记数据的不确定性；其次，为了避免添加随机噪声可能模糊器官边缘的问题，提出一种基于Canny算子的边缘保护噪声以保留图像的边缘信息和重要结构；第三，开发了一种基于均值教师（MT）框架的半监督残差驱动分割方法（RDMT），引入Frobenius范数正则项到指数移动平均方案中，以增强MT的性能。最后，在公开的多器官分割基准数据集BTCV与脑肿瘤分割数据集BraTS 2019上对本文方法进行验证，在BTCV数据集40%有标签数据情况下，Dice相似系数（DSC）与标准化表面距离分别为77.42%与79.47%，在BraTS2019数据集20%有标签数据情况下DSC为83.89%、Jaccard系数为74.21%、平均表面距离为2.34 mm、95%Hausdorff距离为9.08 mm，证明本文方法的优越性。更多还原

【Abstract】 In response to the high cost and time consumption of medical image annotation, and issues such as the imprecision of unlabeled data segmentation in semi-supervised medical image segmentation, loss of image edge information, and delayed parameter updates, a semi-supervised medical image segmentation method based on consistency regularization is presented.Firstly, an uncertainty measurement method based on the dual perspectives of entropy and variance is designed to assess the uncertainty of predictions for unlabeled data, jointly evaluating the uncertainty of unlabeled data from the perspectives of entropy and variance. Then, edge-preserving noise based on the Canny operator is used to retain image edge information and important structures, thereby avoiding the potential blurring of organ edges that may result from the addition of random noise.Finally, a semi-supervised residual-driven segmentation method based on the mean teacher framework is developed, with a Frobenius norm regularization term in the exponential moving average scheme to enhance the performance of mean teacher.The proposed method is validated on the publicly available multi-organ segmentation benchmark dataset BTCV and brain tumor segmentation dataset BraTS 2019. In the case of 40% labeled data in the BTCV dataset, Dice similarity coefficient and standardized surface distance are 77.42% and 79.47%, respectively. In the case of 20% labeled data in the BraTS 2019dataset, the proposed method achieve a Dice similarity coefficient of 83.89%, a Jaccard coefficient of 74.21%, an average surface distance of 2.34 mm, and a 95% Hausdorff distance of 9.08 mm, demonstrating its superiority.更多还原