计算机科学
人工智能
图像分割
模式识别(心理学)
班级(哲学)
分割
计算机视觉
图像(数学)
尺度空间分割
特征提取
图像处理
上下文图像分类
基于分割的对象分类
电子邮件
可视化
医学影像学
图像纹理
作者
Haoxiang Lu,Junjie Yang,Ma Jing,Weidong Zhang,Zhenbing Liu,Zaiyi Liu,Chong Yin
标识
DOI:10.1109/tmm.2026.3739896
摘要
Histopathological image semantic segmentation is crucial for cancer diagnosis and prognosis prediction. Although weakly supervised semantic segmentation (WSSS) methods based on class activation maps (CAM) have reduced reliance on pixel-level annotations, they suffer from incomplete activation, capturing only highly discriminative local features rather than complete semantic regions. This limitation is particularly severe in histopathological images due to tissue co-occurrence and indistinct boundaries, where image-level labels alone provide insufficient supervision for spatial localization. To address this challenge, we propose a histopathological image segmentation method named TP-CAM, which leverages text prompts as global semantic guidance to complement the supervision from imagelevel labels. Unlike image-level labels that only indicate category presence, text embeddings encode holistic semantic concepts that guide the model toward complete region activation. Specifically, we introduce a Text Prompt Gradient Filter (TPGF) module that performs semantic filtering of image features using text guidance in the gradient domain, effectively decoupling discriminative features for co-occurring tissues and localizing distinct regions to generate semantically coherent initial CAMs. Furthermore, we propose a Bidirectional Cross-Attention Refinement (BCAR) module including text-to-image stream and image-to-text stream to balance global semantic guidance with local structural precision for refining initial CAMs. And the former performs semantic verification to ensure activation correctness while the latter incorporates fine-grained visual details to guarantee structural completion. Finally, pseudo-labels with multi-modal semantic information derived from refined CAMs are used to train the segmentation model for histopathological image segmentation. Extensive experiments on two public benchmarks demonstrate that our method achieves average mIoU scores of 0.7764 and 0.7349, outperforming state-of-the-art WSSS methods. Furthermore, the potential benefits of our method to cross-dataset transfer and semantic segmentation for whole slide images are discussed.
科研通智能强力驱动
Strongly Powered by AbleSci AI