Abstract: Deep learning models for histopathological cancer detection often achieve high performance but can rely on non-diagnostic artifacts like staining intensities rather than biological features. This paper introduces M-Saliency, a mathematically rigorous framework that constrains feature extraction to established pathological primitives. By implementing a Unified Morphological Scoring Function, the framework integrates four specific gates: Presence Gate (Tissue identification) Symmetry Ratio (Boundary asymmetry) Boundary Energy (Morphological contrast) Threshold Amplifier (Uncertainty-directed focus) Key Results: The framework was validated across multiple datasets, achieving an AUC of 0.9941 on BreakHis, 0.9953 on LC25000, and 0.9889 on PCam. Crucially, it maintained an AUC of 0.9916 on the independent CRC-VAL-HE-7K validation set, proving its robustness across institutional boundaries. With a modest computational overhead of only 2–3%, M-Saliency offers a practical, interpretable solution for real-time clinical deployment. Keywords: Histopathology, Cancer Detection, Deep Learning, M-Saliency, Morphological Features, Interpretability, Computer Vision, Digital Pathology.
Subhodip Roy (Sun,) studied this question.