摘 要 ：目的 半监督方法旨在通过将未标记数据和标记数据的训练相结合，能够减少对标记数据的依赖，并取得较好的医学图像分割结果。然而，现有的半监督方法通常没有关注到标记数据和未标记数据之间的分布差异所带来的不利影响，尤其是在标记数据比例较低时，可能会严重影响模型的分割性能。方法 提出一种采用分布感知的均值教师网络 (distribution-aware mean teacher network)用于半监督医学影像分割。该方法利用标记数据和未标记数据的分布信息来指导模型的学习，以便在训练阶段使模型对标记和未标记数据的分割结果的分布尽可能相似。该方法采用老师（Teacher）-学生（Student）的网络架构，并嵌入了不同的注意力模块，以及分布感知（distribution-aware，DA）模块、完整性监督（integrity supervision，IS）模块和不确定性最小化（uncertainty minimization，UM）模块。结果 在MICCAI 2018左心房分割挑战LA数据集和胰腺CT数据集上的实验结果表明，该方法使用左心房10%标记数据时，获得的Dice系数、Jaccard指数、HD和ASD分别为88.55%、79.62%、9.07和3.08，与基于不确定性的协同均值教师（UCMT）相比，Dice系数和Jaccard指数分别提高了0.42％和0.44％；而使用左心房20%标记数据时，获得的Dice系数、Jaccard指数、HD距离和ASD距离分别为90.55%、82.82%、5.78和1.77，与UCMT相比，Dice系数和Jaccard指数分别提高了0.14％和0.28％。此外，该方法使用胰腺CT 10%标记数据时，获得的Dice系数、Jaccard指数、HD和ASD分别为70.20%、56.36%、15.64和3.57，与基于不确定性的互补一致性学习（UG-MCL）相比，Dice系数和Jaccard指数分别提高了0.94％和1.06％；而使用胰腺CT 20%标记数据时，获得的Dice系数、Jaccard指数、HD距离和ASD距离分别为77.89%、64.92%、7.97和1.65，与UG-MCL相比，Dice系数和Jaccard指数分别提高了2.77％和3.34％。此外，在ACDC数据集上的实验结果也表明了该方法的优越性。结论 提出的方法利用了标记数据和未标记数据的分布差异信息，有效提升了半监督医学影像分割性能。尤其在使用较低数量的标记数据时，该方法的分割性能明显优于其他使用的半监督方法。

Abstract: Objective Medical image segmentation has great potential for application in clinical diagnosis. While supervised medical image segmentation models can achieve good segmentation results, they heavily rely on pixel-level annotated training data. However, acquiring labeled data is costly and can only be done by experts. Especially in the case of 3D medical datasets, the process of annotation is more intricate and time-consuming compared to 2D datasets. Semi-supervised learning is an effective solution that combines labeled and unlabeled data for improved segmentation results. However, existing semi-supervised methods have not addressed the performance impact caused by the distribution gap between labeled and unlabeled data, especially when the proportion of labeled data is low. Method To reduce the performance impact caused by the distribution gap between labeled and unlabeled data, a distribution-aware mean teacher network (DAMTN) is proposed for semi-supervised medical image segmentation. This method utilizes the distribution information of both labeled and unlabeled data to guide the learning of the model, aiming to make the segmentation results of labeled and unlabeled data have similar distributions during the training phase. The DAMTN adopts a Teacher-Student architecture, consisting of a teacher model and a student model. In addition, both the teacher model and the student model"s network architectures are based on the V-Net design, which includes an encoder and three decoders. The residual connections between the encoder and decoder have been removed. The decoders are differentiated by embedding different attention modules into them, which are used to introduce perturbations at the model level. These attentions, including Cross-sample Mutual Attention (CMA), Position Attention (PA), and Channel Attention (CA), are employed to process high-level features. CMA is used for inter-sample feature interaction and alignment, PA is employed to handle spatial position information in the feature maps, and CA is utilized to address channel correlations in the feature maps. And DAMTN comprises three key modules: the distribution-aware (DA) module, integrity supervision (IS) and uncertainty minimization (UM) module. These three modules compute the losses for the outputs of the student model, and the outputs of the teacher model are used to calculate the consistency loss with the corresponding outputs of the student model. The DA module encourages the model to learn features when the distribution information from labeled and unlabeled data is similar. To obtain the distribution information of labeled and unlabeled data, each decoder branch is equipped with a dual normalization block, consisting of two normalization layers, referred to as and . handles the unlabeled data, while handles the labeled data. The IS module supervises the predicted mean of the branches and encourages the model to focus on challenging regions, further reducing the differences between branches. The UM module aims to balance the uncertainty of different attention branches, thereby making the predictions of all branches more consistent and improving the model"s confidence. To improve the segmentation results, the student model incorporates an integrity supervision module and an uncertainty minimization module during the training phase, which ensures the consistency and accuracy of the outputs from the three decoders. During the testing phase, the student model takes the average output of the three decoders as its final output. In addition, the parameters of the student model are updated by the optimizer, while the parameters of the teacher model are updated through exponential moving average (EMA). Result Experiments were conducted on the MICCAI 2018 (Medical Image Computing and Computer Assisted Intervention Society) Left Atrial Segmentation Challenge dataset and the Pancreas CT dataset. In the experiment with 10% labeled data for left atrium segmentation, the Dice coefficient, Jaccard index, hausdorff distance (HD), and average surface distance (ASD) were 88.55%, 79.62%, 9.07, and 3.08, respectively. In the experiment with 20% labeled data for left atrium, the Dice coefficient, Jaccard index, HD, and ASD were 90.55%, 82.82%, 5.78, and 1.77, respectively. On the LA dataset, when trained with 10% labeled data, compared to uncertainty-guided collaborative mean teacher (UCMT), the Dice coefficient and Jaccard index improved by 0.42% and 0.44% respectively. On the LA dataset, when trained with 20% labeled data, compared to UCMT, the Dice coefficient and Jaccard index improved by 0.14% and 0.28% respectively. In the experiment with 10% labeled data for Pancreas CT, the Dice coefficient, Jaccard index, HD, and ASD were 70.20%, 56.36%, 15.64, and 3.57 respectively. In the experiment with 20% labeled data for Pancreas CT, the Dice coefficient, Jaccard index, HD, and ASD were 77.89%, 64.92%, 7.97, and 1.65 respectively. On the Pancreas dataset, when trained with 10% labeled data, compared to uncertainty-guided mutual consistency learning (UG-MCL), the Dice coefficient and Jaccard index improved by 0.94% and 1.06% respectively. Additionally, compared to mutual consistency network (MC-Net), the Dice coefficient and Jaccard index improved by 3.70% and 4.00% respectively. On the Pancreas dataset, when trained with 20% labeled data, compared to UG-MCL, the Dice coefficient and Jaccard index improved by 2.77% and 3.34% respectively. Additionally, compared to MC-Net, the Dice coefficient and Jaccard index improved by 0.73% and 0.61% respectively. Additionally, the experimental results on the ACDC dataset also demonstrate the superiority of this method. Conclusion The proposed model in this paper is based on the teacher-student framework, incorporating attention mechanisms and leveraging the distribution information of both labeled and unlabeled data to constrain the student model. It effectively addresses the distribution gap between labeled and unlabeled data, reducing its impact on performance and improving segmentation results. Particularly, when the labeled data is scarce, this model outperforms other semi-supervised segmentation methods in terms of segmentation performance.