ICAM: An interpretable auxiliary model for the pathological diagnosis of breast cancer based on knowledge embedding

Pathological diagnosis is the gold standard for cancer diagnosis. To provide reliable pathological diagnoses, this study proposes a pathological image diagnostic model based on knowledge embeddings. The model effectively guides the learning of pathological diagnostic features by embedding domain knowledge, thereby improving its efficiency. On this basis, nonnegative matrix factorization (NMF) is applied locally. As a powerful feature decomposition method, NMF can effectively extract local patterns and texture information from images, enhancing the capability of representing features and further strengthening the model's ability to capture key features. Finally, the decision basis of the model is derived through backward deduction of the model's computational steps, providing clear diagnostic reasoning for clinical doctors. The validation results on both private and public datasets show that, compared with conventional models, the proposed approach improves the accuracy (ACC) by 15.8% and the F1 score by 15.4%. The comparison with other methods highlights the robustness of ICAM and underscores its potential advantages over other transformer-based approaches. Additionally, the model's decision-making rationale is revealed through clear visual explanations, which are consistent with clinical observations, demonstrating its practical utility and clinical reference value.