增强不确定性引导均值教师模型的HER2组织病理图像半监督分割模型

Enhanced Uncertainty-Guided Mean Teacher Model for Semi-Supervised Segmentation Model of HER2 Histopathological Images

摘要: 在乳腺癌诊断中，为了解决大量需求HER2组织病理染色图像专家标注的问题，传统半监督模型会过早收敛为自训练，导致生成低置信度为标签。为了应对这些挑战，本文提出一种新的增强不确定性引导的半监督分割网络，用于HER2组织病理染色图像的细胞膜分割。首先，该网络在协同均值教师模型中引入一个不确定目标选择模块，该模块通过构建联合置信度矩阵，利用学生模型和教师模型之间的预测差异，有效识别出由于系统性误差导致的伪标签噪声区域。其次，将识别出的不确定区域加入到一致性损失中，引导教师学生模型在不确定目标区域预测中达成共识，增强模型像素级目标选择能力。此外，该方法无需修改主干网络，也不增加额外的可训练参数，具备良好的泛化能力和可扩展性。我们的实验在HER2组织病理染色图片数据集中得到验证，结果表明，我们的方法在Dice评价指标下得到73.09%的分数，相较于基线模型有明显提升，同时我们的方法也超越了最新的SSL方法，也优于经典的全监督学习方法。该方法不仅在HER2组织病理染色图片数据集分割任务中表现优异，在宫颈癌细胞核分割任务中Dice评价指标也达到85.02%，也具备作为半自动标注工具在其他疾病领域的广泛应用。

Abstract: In breast cancer diagnosis, the extensive demand for expert annotations of HER2-stained histopathological images poses a major bottleneck. Conventional semi-supervised learning frameworks tend to converge prematurely into self-training, which often leads to the generation of low-confidence pseudo-labels. To address these challenges, we propose a novel enhanced uncertainty-guided semi-supervised segmentation network for cell membrane segmentation in HER2-stained histopathological images. Specifically, an uncertainty-aware target selection module is introduced into a collaborative mean teacher framework. This module constructs a joint confidence matrix and exploits the prediction discrepancies between the student and teacher models to effectively identify pseudo-label noisy regions caused by systematic errors. Subsequently, the identified uncertain regions are incorporated into the consistency loss, guiding the teacher-student models to reach consensus on uncertain targets and thereby enhancing pixel-level target selection capability. Notably, the proposed method does not require any modification of the backbone network nor introduce additional trainable parameters, demonstrating strong generalization and scalability. Extensive experiments conducted on the HER2-stained histopathological image dataset validate the effectiveness of the proposed approach. Our method achieves a Dice score of 73.09%, showing a substantial improvement over the baseline model, while also outperforming recent state-of-the-art semi-supervised learning methods and even classical fully supervised approaches. Furthermore, the proposed framework exhibits strong generalization performance on cervical cancer nucleus segmentation, attaining a Dice score of 85.02%. These results indicate that our method is not only effective for HER2 histopathological image segmentation but also holds significant potential as a semi-automatic annotation tool for a wide range of diseases.