Colorectal cancer is a common and lethal malignant tumor, where histopathological examination serves as the gold standard for diagnosis. With the rise of digital pathology, automated analysis of histopathological images has become crucial. However, the high cost and expertise required for pixel-level annotations present a significant bottleneck. Weakly supervised semantic segmentation (WSSS), which utilizes image-level labels, offers a promising and cost-effective alternative. Nevertheless, existing WSSS methods applied to pathology images often suffer from two critical issues: they generate incomplete class activation maps (CAMs) that only highlight the most discriminative parts of a lesion, and they are susceptible to interference from complex contextual noise. To address these challenges, we propose CPPD-DMamba, a dual-phase enhancement framework based on a dynamic selective state space model that cooperates with contextual prototypes and point diffusion to generate complete and consistent segmentation results. Specifically, we construct a dynamic visual Mamba (DVM) with geometric deformation to accommodate the significant size variations and diverse morphological characteristics of histopathological cells. Furthermore, we design a contextual prototype alignment module that employs clustering strategies to filter category-specific contextual features, thereby mitigating interference from irrelevant contextual information. Finally, we propose a key point diffusion module to enhance boundary integrity by expanding target regions. Comprehensive experiments on three public pathology image datasets (CRAG, EBHI-SEG, and GLAS) demonstrate that our CPPD-DMamba significantly outperforms state-of-the-art methods, validating its effectiveness and potential for clinical application.
Wang et al. (Thu,) studied this question.