Feature Contribution Analysis Across Multimodal Breast Cancer Diagnostic Models
Résumé
This study systematically assesses and contrast the interpretability and clinical usefulness of SHAP-based explanations in multimodal breast cancer diagnosis models. A cross-task meta-analysis was conducted, aggregating the results of five state-of-the-art multimodal breast cancer diagnostic studies addressing benign-malignant differentiation, HER2 expression status, axillary lymph node metastasis, Ki-67 proliferation assessment and molecular subtype classification. Outputs from the key studies across the four modalities of ultrasound, mammography, MRI, PET/CT and structured clinical data were harmonized to the same feature families and compared using global feature importance rankings and case-level analysis. The synthesis showed that peritumoral texture entropy and skewness were the most significant variables for benign-malignant differentiation and PET/CT metabolic parameters and tumor diameter for predicting HER2 status with an AUC of 0.888 training and an AUC of 0.844 test. A mammographic descriptor, indistinct margins, and a tumor micro-textural heterogeneity were significant for predicting axillary lymph node metastasis and Ki-67 status, while a mammographic descriptor, calcification, drove molecular subtyping. Case-level SHAP explanations revealed error mechanisms such as a false negative metastasis case due to over-reliance on tumour shape features. Radiologists with the help of models that included SHAP, gained 8.3-15.9 percentage points more accuracy in the subtype classification, with the highest improvement coming from less experienced readers. This study demonstrates that the SHAP-based interpretability approach is a feasible tool for human-in-the-loop transparency and auditability of precision breast oncology decision-making, which allows clinicians to gain insight into model reasoning, detect model failure points, and enhance diagnostic accuracy in various multimodal diagnostic tasks.
Citer ce document
Exporter : BibTeX · RIS (Zotero, Mendeley, EndNote)
Accès au document
Texte intégral archivé sur SPHAERO, en accès ouvert : lisible gratuitement en ligne après connexion. Le document reste aussi disponible à sa source. Se connecter
Licence et provenance
Licence : CC BY-SA
Notice moissonnée depuis OpenAlex le 26/10/2026. Le document reste hébergé par sa source.
Voir le document à la source →