Co-training semi-supervised medical image segmentation based on pseudo-label weight balancing.

Purpose: Major challenges in current semi-supervised segmentation methods: (1) The complementary nature of information in pseudo-label: a key limitation of consistent regularization methods is the tendency of sub-networks to converge to the consensus case early on, leading to the degradation of co-trained models into self-trained models, whereas disagreement between sub-networks is important for joint training. (2) Quantity-quality weighting imbalance in pseudo-label methods: threshold-based pseudo-label is to train the model with pseudo-labels whose predicted confidence is higher than a hard threshold. In contrast, other pseudo-labels are simply ignored. This study aims to propose a semi-supervised model based on pseudo-labeled weight balancing for medical image segmentation tasks for the above-mentioned problems.

Methods: We adopted a truncated Gaussian function weight balancing method based on the marginal hypothesis distribution to generate high-quality pseudo-labels while maintaining a high utilization rate of pseudo-labels, and based on which we applied a uniform alignment strategy to solve the pseudo-label imbalance problem due to the difference in the learning difficulty of different classes. In addition, to address the problem that self-training algorithms rely too much on the quality of pseudo-labels generated, we inherit the idea of knowledge refinement and integrate the mean teacher model of co-training, thus proposing a novel semi-supervised medical image segmentation framework, SCMT (Semi-supervised Co-training Mean Teacher), which is aimed at improving the existing self-training algorithms or co-training algorithms limitations of a single model.

Results: We validate the effectiveness of the method by performing experimental evaluation on two commonly used benchmark medical datasets, LA, and Pancreas-CT, by using 10%/20% labeled data and 90%/80% unlabeled data for training. On the LA dataset, the model obtained Hausdorff distance (HD) of 6.65 mm/5.63 mm, average symmetric surface distance of 1.91 mm/0.02 mm, Dice similarity coeffcient of 90.09%/91.05%, and Jaccard of 81.08%/83.64%. On the Pancreas-CT dataset, the model obtained HD of 12.71 mm/6.63 mm, average symmetric surface distance of 2.01 mm/1.27 mm, Dice similarity coeffcient of 74.64%/81.77% and Jaccard of 60.48%/69.51%. The results show that our method not only outperforms existing semi-supervised segmentation methods but also significantly improves segmentation performance and reduces the dependence on labeled data to achieve consistent and stable prediction results.

Conclusions: We proposed a weight-balanced co-trained cross-consistent semi-supervised model SCMT for semi-supervised segmentation of medical images, which consists of a CMT (Co-training Mean Teacher) structure and quantity-quality-balanced pseudo-label-guided mutual consistency constraints. Compared with other models, we effectively exploit the challenging region and can more accurately capture the contours and finer details of the segmented objects without any shape or boundary constraints, resulting in highly accurate and detail-rich segmentation results. In addition, we conduct comparative experiments with existing semi-supervised models, and the experimental results show that our proposed model is capable of handling complex structures and segmenting details commonly missed by other methods. The segmentation results obtained are relatively stable and consistent and have certain advantages in improving the performance of surface segmentation. Code is available at: https://github.com/zhaojiashi/SCMT.