MCCNet: A Multi-granularity Collaborative Cascaded-perception Network for Colorectal Cancer Histopathological...

Colorectal cancer (CRC) is among the most lethal malignancies worldwide, and histopathological examination remains the clinical gold standard for diagnosis. In the histopathological examination, collaborative perception of fine-grained and coarse-grained lesions in histopathological samples is crucial for CRC diagnosis. However, existing methods often struggle to simultaneously model discriminative cues across granularities because of fine-grained glandular disorganization and coarse-grained morphological overlap among tissue categories. To tackle these issues, we propose a novel Multi-granularity Collaborative Cascaded-perception Network (MCCNet), which learns cross-granularity discriminative representations to overcome the limitations of single-granularity modeling. For fine-grained glandular disorganization, we design an Adaptive Glandular Offset Convolution (AGO-Conv). The convolution adaptively learns sampling offsets by injecting prominent edge gradients to actively focus on disorganized glands, and dynamically adjusts the sampling radius by a continuous domain constraint mechanism, thereby preventing out-of-bound glandular sampling. For coarse-grained similar lesions, we construct a Grouped Residual Linear-Attention (GRLA) module. GRLA linear approximate exponential kernel by segmented decomposable kernel mapping, and the discriminative inter-class representations across multiple subspaces are further preserved by subspace grouping and deep residual paths, thereby alleviating inter-class similarity among different tissue categories. To enable dynamic cross-granularity feature coupling, a Multi-Granularity Cascaded Feature Fusion (MGCFF) module is proposed that recalibrates the discriminative responses along channel and spatial dimensions via cascaded hierarchical representations and feature recalibration units, and further eliminates low-response background noise by multi-scale gating mechanism, thereby mitigating cross-granularity semantic conflicts. Extensive experimental results on the Chaoyang, Kather-5K, EBHI, and Private dataset show that MCCNet consistently outperforms existing methods.