National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging, School of Biomedical Engineering, Health Science Center, Shenzhen University, Shenzhen, China.
Smart Medical Imaging, Learning and Engineering (SMILE) Lab, Shenzhen University, Shenzhen, China.
Med Phys. 2023 Jan;50(1):163-177. doi: 10.1002/mp.15923. Epub 2022 Aug 23.
Semisupervised strategy has been utilized to alleviate issues from segmentation applications due to challenges in collecting abundant annotated segmentation masks, which is an essential prerequisite for training high-performance 3D convolutional neural networks (CNNs) .
Existing semisupervised segmentation methods are mainly concerned with how to generate the pseudo labels with regularization but not evaluate the quality of the pseudo labels explicitly. To alleviate this problem, we offer a simple yet effective reciprocal learning strategy for semisupervised volumetric medical image segmentation, which generates more reliable pseudo labels for the unannotated data.
Our proposed reciprocal learning is achieved through a pair of networks, one as a teacher network and the other as a student network. The student network learns from pseudo labels generated by the teacher network. In addition, the teacher network autonomously optimizes its parameters based on the reciprocal feedback signals from the student's performance on the annotated images. The efficacy of the proposed method is evaluated on three medical image data sets, including 82 pancreas computed tomography (CT) scans (training/testing: 62/20), 100 left atrium gadolinium-enhanced magnetic resonance (MR) scans (training/testing: 80/20), and 200 breast cancer MR scans (training/testing: 68/132). The comparison methods include mean teacher (MT) model, uncertainty-aware MT (UA-MT) model, shape-aware adversarial network (SASSNet), and transformation-consistent self-ensembling model (TCSM). The evaluation metrics are Dice similarity coefficient (Dice), Jaccard index (Jaccard), 95% Hausdorff distance (95HD), and average surface distance (ASD). The Wilcoxon signed-rank test is used to conduct the statistical analyses.
By utilizing 20% labeled data and 80% unlabeled data for training, our proposed method achieves an average Dice of 84.77%/90.46%/78.53%, Jaccard of 73.71%/82.67%/69.00%, ASD of 1.58/1.90/0.57, and 95HD of 6.24/5.97/4.34 on pancreas/left atrium/breast data sets, respectively. These results outperform several cutting-edge semisupervised approaches, showing the feasibility of our method for the challenging semisupervised segmentation applications.
The proposed reciprocal learning strategy is a general semisupervised solution and has the potential to be applied for other 3D segmentation tasks.
由于在收集大量标注分割掩模方面存在挑战，半监督策略已被用于缓解分割应用中的问题，这是训练高性能 3D 卷积神经网络 (CNN) 的必要前提。
现有的半监督分割方法主要关注如何通过正则化生成伪标签，而没有显式评估伪标签的质量。为了解决这个问题，我们为半监督容积医学图像分割提供了一种简单而有效的互惠学习策略，为未标注数据生成更可靠的伪标签。
我们提出的互惠学习是通过一对网络实现的，一个是教师网络，另一个是学生网络。学生网络从教师网络生成的伪标签中学习。此外，教师网络还根据学生在标注图像上的性能的反馈信号，自主优化其参数。在三个医学图像数据集上评估了所提出方法的效果，包括 82 个胰腺 CT 扫描（训练/测试：62/20）、100 个左心房钆增强磁共振 (MR) 扫描（训练/测试：80/20）和 200 个乳腺癌 MR 扫描（训练/测试：68/132）。比较方法包括平均教师 (MT) 模型、不确定性感知 MT (UA-MT) 模型、形状感知对抗网络 (SASSNet) 和变换一致自集成模型 (TCSM)。评估指标包括 Dice 相似系数 (Dice)、Jaccard 指数 (Jaccard)、95%Hausdorff 距离 (95HD) 和平均表面距离 (ASD)。采用 Wilcoxon 符号秩检验进行统计分析。
利用 20%的标注数据和 80%的未标注数据进行训练，我们提出的方法在胰腺/左心房/乳房数据集上的平均 Dice 分别为 84.77%/90.46%/78.53%、Jaccard 分别为 73.71%/82.67%/69.00%、ASD 分别为 1.58/1.90/0.57 和 95HD 分别为 6.24/5.97/4.34。这些结果优于几种前沿的半监督方法，表明我们的方法在具有挑战性的半监督分割应用中是可行的。
所提出的互惠学习策略是一种通用的半监督解决方案，有可能应用于其他 3D 分割任务。
