[Colorectal cancer diagnosis method based on dynamic gland-aware and tissue soft-clustering]

In colorectal cancer histopathology, the collaborative perception of microscopic and salient lesions is critical for effective diagnosis and improved patient prognosis. However, existing deep learning methods struggle to simultaneously capture microscopic glandular disorganization and salient tissue lesions. To address this limitation, a colorectal cancer diagnosis network based on dynamic gland-aware and tissue soft-clustering (DGTSNet) is proposed. The method employs dynamic gland-aware convolution to explicitly perceive gland boundaries and dynamically adjust sampling offsets, while incorporating a continuous-domain constraint to prevent out-of-bound sampling, thereby enhancing the perception of microscopic glandular disorders. Meanwhile, a tissue soft-clustering module is utilized to adaptively generate clustering prototypes and guide pixels toward relevant prototype centroids according to semantic similarity, suppressing irrelevant background interference and enhancing responses to significant tissue lesions. Finally, a cascaded sparse coupling module is introduced to collaboratively modulate cross-semantic feature representations along both the spatial and channel dimensions, while constructing differentiable masks to suppress low-contribution semantics, thereby achieving collaborative coupling of cross-semantic features. Experimental results demonstrated that the proposed method achieved superior performance on the Chaoyang, Kather-5K, and EBHI datasets, as well as a real-world clinical validation cohort, outperforming multiple baseline models. The study shows that the proposed method can effectively enhance the joint perception of microscopic glandular disorders and significant tissue lesions, providing an effective solution for colorectal cancer diagnosis.