摘要：

医学图像分割在疾病辅助诊断中起着关键的作用。现有的深度分割模型需要依赖带有标注的数据完成大规模训练, 而医学影像标注需要具有专业背景的临床医生进行像素级标注, 导致标注数据获取困难。基于半监督的医学图像分割方法利用少量的标注数据和大量的未标注数据进行学习, 可以在一定程度上缓解标注数据获取困难的问题。针对半监督分割模型不能充分利用未标注数据中的可学习信息的问题, 提出一种半监督分割模型TCA-Net。该模型使用U-Net作为骨干网络, 通过在U-Net中引入卷积块注意力模块(CBAM)与多头自注意力模块(MHA)来解决其在下采样过程中的信息丢失问题; 为了充分利用未标注数据中的不确定性信息, 构建一个教师互一致性模型, 该模型由具有1个编码器和3个略有不同的解码器的学生模型与教师模型组成, 通过在学生模型的概率映射与教师模型的伪标签之间添加一致性约束, 以此在训练过程中最小化输出之间的差异, 从而提升模型的分割效果。在公开的WORD腹部多器官数据集与ACDC心脏数据集上进行实验, 结果表明, 在使用20%标注数据的WORD数据集上, TCA-Net的Dice系数、Jaccard指数、HD95和ASD分别达到90.81%、83.79%、21.38和6.08, 在ACDC数据集上分别达到89.69%、81.94%、1.66和0.45。消融实验与对比实验结果表明, TCA-Net能够有效提升未标注数据的利用率, 在不同数据集上均达到了较好的分割效果, 验证了模型的鲁棒性。

关键词: 医学图像分割, 半监督学习, 注意力机制, 平均教师模型, 一致性正则化

Abstract:

Medical image segmentation plays a crucial role in disease-assisted diagnosis. Existing deep segmentation models rely on annotated data for large-scale training. Medical image annotation requires clinical doctors with professional backgrounds to perform pixel-level annotation, making it difficult to obtain annotation data. The semi-supervised medical image segmentation method utilizes a small amount of labeled data and a large amount of unlabeled data for learning, which can alleviate the difficulty of obtaining labeled data completely. This study proposes a semi-supervised segmentation network, TCA-Net to address the issue of semi-supervised segmentation models that do not fully utilize learnable information from unlabeled data. This network uses U-Net as the backbone network. This network solves the problem of information loss during downsampling by introducing a Convolutional Block Attention Module (CBAM) and Multi-Head self Attention module (MHA) in the U-Net network. To utilize the uncertainty information in unlabeled data fully, this network employs the construction of a teacher mutual consistency model. The model consists of a student model with an encoder and three slightly different decoders and a teacher model. By adding consistency constraints between the probability mapping of the student model and the pseudo labels of the teacher model, this network succeeds in minimizing the difference between outputs during training, thereby improving the segmentation effect of the model. To verify the effectiveness of the proposed model, experiments are conducted on the publicly available WORD abdominal multi-organ and ACDC heart datasets. On the WORD dataset using 20% labeled data, the Dice coefficient, Jaccard index, HD95, and ASD reach 90.81%, 83.79%, 21.38, and 6.08, respectively, whereas they reach 89.69%, 81.94%, 1.66, and 0.45, respectively, on the ACDC dataset. By comparing the results of ablation experiments with those of advanced algorithms, TCA-Net effectively improves the utilization of unlabeled data and achieves good segmentation results on different datasets, verifying the robustness of the model.

Key words: medical image segmentation, semi-supervised learning, attention mechanism, average teacher model, consistent regularization