Xue Yuanchao, Hu Yangsheng, Yao Yu, Huang Jie, Wang Haitao, He Jianfeng
Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650504, China.
Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650504, China.
Med Eng Phys. 2025 Feb;136:104284. doi: 10.1016/j.medengphy.2025.104284. Epub 2025 Jan 6.
Accurate semantic segmentation of histopathological images plays a crucial role in accurate cancer diagnosis. While fully supervised learning models have shown outstanding performance in this field, the annotation cost is extremely high. Weakly Supervised Semantic Segmentation (WSSS) reduces annotation costs due to the use of image-level labels. However, these WSSS models that rely on Class Activation Maps (CAM) focus only on the most salient parts of the image, which is challenging when dealing with semantic segmentation tasks involving multiple targets. We propose a two-stage weakly supervised segmentation framework (MSRMMP) to resolve the above problems, the generation of pseudo masks based on multi-scale residual networks (MSR-Net) and the semantic segmentation based on multi-layer pseudo-supervision. MSR-Net fully captures the local features of an image through multi-scale residual module (MSRM) and generates pseudo masks using image-level label. Additionally, we employ Transunet as the segmentation backbone, and uses multi-layer pseudo-supervision algorithms to solve the problem of pseudo-mask inaccuracy. Experiments performed on two publicly available histopathology image datasets show that our proposed method outperforms other state-of-the-art weakly supervised semantic segmentation methods. Additionally, it outperforms the fully-supervised model in mIoU and has a similar result in fwIoU when compared to fully-supervised models. Compared with manual labeling, our model can significantly save the labeling time from hours to minutes.
组织病理学图像的精确语义分割在准确的癌症诊断中起着至关重要的作用。虽然全监督学习模型在该领域表现出色，但标注成本极高。弱监督语义分割（WSSS）由于使用图像级标签而降低了标注成本。然而，这些依赖类激活映射（CAM）的WSSS模型仅关注图像中最显著的部分，在处理涉及多个目标的语义分割任务时具有挑战性。我们提出了一个两阶段的弱监督分割框架（MSRMMP）来解决上述问题，即基于多尺度残差网络（MSR-Net）生成伪掩码以及基于多层伪监督进行语义分割。MSR-Net通过多尺度残差模块（MSRM）充分捕捉图像的局部特征，并使用图像级标签生成伪掩码。此外，我们采用Transunet作为分割主干，并使用多层伪监督算法来解决伪掩码不准确的问题。在两个公开可用的组织病理学图像数据集上进行的实验表明，我们提出的方法优于其他现有的弱监督语义分割方法。此外，如果与全监督模型相比，它在平均交并比（mIoU）方面优于全监督模型，在频率加权交并比（fwIoU）方面有类似的结果。与手动标注相比，我们的模型可以将标注时间从数小时显著缩短至数分钟。
