Digital pathology poses challenges due to stain variability, complex tissue forms, or heterogeneous conditions. Conventional deep models operate directly on raw RGB inputs. Unlike them, the proposed model—which is a cognitively inspired and lightweight one—integrates domain-aware preprocessing. These include stain normalization, contrast enhancement, edge detection, and discretization. After these principles, tokenization and feature mixing are applied. The core architecture employs mixer blocks with dynamic CheapOp gating to selectively fuse transformations. Extensive experiments on an eight-class colorectal histology dataset show that this model outperforms CNN, transformer, and other mixer-based baselines. Ablation further confirms the contribution of each preprocessing and architectural component. It highlights the robustness of the pipeline and, potentially, deployment suitability.