Yang Fan, He Qiming, Wang Yanxia, Zeng Siqi, Xu Yingming, Ye Jing, He Yonghong, Guan Tian, Wang Zhe, Li Jing
Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Shenzhen, China.
Department of Pathology, State Key Laboratory of Cancer Biology, Xijing Hospital, Fourth Military Medical University, Xi'an, China.
Int J Comput Assist Radiol Surg. 2025 Feb;20(2):225-236. doi: 10.1007/s11548-024-03154-7. Epub 2024 Jun 7.
In pathology images, different stains highlight different glomerular structures, so a supervised deep learning-based glomerular instance segmentation model trained on individual stains performs poorly on other stains. However, it is difficult to obtain a training set with multiple stains because the labeling of pathology images is very time-consuming and tedious. Therefore, in this paper, we proposed an unsupervised stain augmentation-based method for segmentation of glomerular instances.
In this study, we successfully realized the conversion between different staining methods such as PAS, MT and PASM by contrastive unpaired translation (CUT), thus improving the staining diversity of the training set. Moreover, we replaced the backbone of mask R-CNN with swin transformer to further improve the efficiency of feature extraction and thus achieve better performance in instance segmentation task.
To validate the method presented in this paper, we constructed a dataset from 216 WSIs of the three stains in this study. After conducting in-depth experiments, we verified that the instance segmentation method based on stain augmentation outperforms existing methods across all metrics for PAS, PASM, and MT stains. Furthermore, ablation experiments are performed in this paper to further demonstrate the effectiveness of the proposed module.
This study successfully demonstrated the potential of unsupervised stain augmentation to improve glomerular segmentation in pathology analysis. Future research could extend this approach to other complex segmentation tasks in the pathology image domain to further explore the potential of applying stain augmentation techniques in different domains of pathology image analysis.
在病理图像中，不同的染色方法会突出显示不同的肾小球结构，因此基于监督深度学习的肾小球实例分割模型在单一染色上训练后，对其他染色的表现较差。然而，由于病理图像的标注非常耗时且繁琐，很难获得包含多种染色的训练集。因此，在本文中，我们提出了一种基于无监督染色增强的肾小球实例分割方法。
在本研究中，我们通过对比无配对翻译（CUT）成功实现了诸如PAS、MT和PASM等不同染色方法之间的转换，从而提高了训练集的染色多样性。此外，我们用Swin Transformer替换了Mask R-CNN的主干，以进一步提高特征提取效率，从而在实例分割任务中取得更好的性能。
为了验证本文提出的方法，我们从本研究中的216张三种染色的全切片图像（WSIs）构建了一个数据集。经过深入实验，我们验证了基于染色增强的实例分割方法在PAS、PASM和MT染色的所有指标上均优于现有方法。此外，本文还进行了消融实验，以进一步证明所提出模块的有效性。
本研究成功证明了无监督染色增强在病理分析中改善肾小球分割的潜力。未来的研究可以将这种方法扩展到病理图像领域的其他复杂分割任务，以进一步探索在病理图像分析的不同领域应用染色增强技术的潜力。
