CSNet: Cognitive Simulation Network for Reliable Histopathology-Based Cancer Diagnosis.

Histopathology-based cancer diagnosis is crucial for early detection and timely therapeutic intervention. However, most deep learning methods for histopathological image classification are mainly optimized as prediction models, with limited consideration of clinical diagnostic workflows. In routine practice, pathologists usually follow a sequential reasoning process: coarse assessment of possible abnormalities, detailed inspection of suspicious regions, and consultation-based reasoning to reduce uncertainty. To provide more reliable and interpretable decision support for histopathology-based cancer diagnosis, we propose the cognitive simulation network (CSNet), which draws inspiration from pathologists' sequential diagnostic workflow. First, the multi-view fuzzy assessment module integrates frequency, texture, and gradient cues through learnable fuzzy membership functions to estimate view-level abnormality degrees, providing an initial assessment of suspicious pathological evidence. Guided by this assessment, the gradient-guided fuzzy modulator converts abnormality information into spatial guidance, enhancing informative high-gradient morphological responses while suppressing redundant low-informative regions. Based on the modulated representations, the uncertainty-guided multi-expert synergy mechanism estimates diagnostic uncertainty and coordinates multiple expert branches to generate consensus representations for reliable prediction. Extensive experiments on Kather-5 K, LC25000, LC25000-clean, and a collected cervical cancer histopathology dataset demonstrate that CSNet achieves superior performance, with accuracies of 97.673%, 99.964%, 97.920%, and 99.861%, respectively. Ablation studies validate the effectiveness of the proposed designs, and statistical analyses confirm the reliability of the observed improvements. Meanwhile, robustness evaluations and comparative visualizations demonstrate stable performance and more consistent responses in diagnostically relevant tissue regions, highlighting the potential of CSNet for reliable histopathology-based cancer diagnosis.