PROSPECTIVE MULTICENTER VALIDATION OF EXPLAINABLE ARTIFICIAL INTELLIGENCE–INTEGRATED RADIOMICS FOR PRECISION DIAGNOSIS AND OUTCOME PREDICTION IN ONCOLOGIC IMAGING
ABSTRACT
Background: Radiomics and artificial intelligence (AI) have emerged as promising tools for precision oncology by enabling quantitative characterization of tumor heterogeneity from medical images. However, limited interpretability and insufficient prospective multicenter validation remain major barriers to clinical adoption. This study aimed to evaluate the diagnostic and prognostic performance of an explainable artificial intelligence (XAI)-integrated radiomics framework in oncologic imaging.
Materials and Methods: In this prospective multicenter study, 612 patients with histopathologically confirmed solid malignancies from five tertiary oncology centers were enrolled between January 2023 and January 2026. Radiomic features were extracted from standardized CT, MRI, and PET/CT examinations and incorporated into machine-learning models. Explainability was achieved using SHapley Additive exPlanations (SHAP). Model performance was assessed using area under the receiver operating characteristic curve (AUC), calibration metrics, decision-curve analysis, and survival modeling. Independent internal and external validation cohorts were used to evaluate generalizability.
Result: The XAI-radiomics model demonstrated superior diagnostic performance compared with clinical, conventional radiomics, and deep learning models, achieving an external validation AUC of 0.921 (95% CI: 0.896–0.944), sensitivity of 89.7%, specificity of 86.8%, and overall accuracy of 88.2%. SHAP analysis identified entropy, gray-level non-uniformity, tumor volume, and wavelet-derived texture features as the most influential predictors. During a median follow-up of 28 months, the XAI-radiomics risk score independently predicted overall survival (HR: 3.42, 95% CI: 2.46–4.75; p<0.001). The model achieved a C-index of 0.83 and demonstrated excellent calibration and clinical utility across validation cohorts.
Conclusion: Explainable AI-integrated radiomics provides robust, interpretable, and generalizable diagnostic and prognostic assessment in oncologic imaging. The proposed framework has substantial potential to support precision oncology and facilitate the clinical translation of transparent AI-driven imaging biomarkers.
Keywords: Explainable Artificial Intelligence, Radiomics, Precision Oncology, Oncologic Imaging, Machine Learning, SHAP, Imaging Biomarkers, Prognostication, Survival Prediction, Artificial Intelligence.
Background: Radiomics and artificial intelligence (AI) have emerged as promising tools for precision oncology by enabling quantitative characterization of tumor heterogeneity from medical images. However, limited interpretability and insufficient prospective multicenter validation remain major barriers to clinical adoption. This study aimed to evaluate the diagnostic and prognostic performance of an explainable artificial intelligence (XAI)-integrated radiomics framework in oncologic imaging.
Materials and Methods: In this prospective multicenter study, 612 patients with histopathologically confirmed solid malignancies from five tertiary oncology centers were enrolled between January 2023 and January 2026. Radiomic features were extracted from standardized CT, MRI, and PET/CT examinations and incorporated into machine-learning models. Explainability was achieved using SHapley Additive exPlanations (SHAP). Model performance was assessed using area under the receiver operating characteristic curve (AUC), calibration metrics, decision-curve analysis, and survival modeling. Independent internal and external validation cohorts were used to evaluate generalizability.
Result: The XAI-radiomics model demonstrated superior diagnostic performance compared with clinical, conventional radiomics, and deep learning models, achieving an external validation AUC of 0.921 (95% CI: 0.896–0.944), sensitivity of 89.7%, specificity of 86.8%, and overall accuracy of 88.2%. SHAP analysis identified entropy, gray-level non-uniformity, tumor volume, and wavelet-derived texture features as the most influential predictors. During a median follow-up of 28 months, the XAI-radiomics risk score independently predicted overall survival (HR: 3.42, 95% CI: 2.46–4.75; p<0.001). The model achieved a C-index of 0.83 and demonstrated excellent calibration and clinical utility across validation cohorts.
Conclusion: Explainable AI-integrated radiomics provides robust, interpretable, and generalizable diagnostic and prognostic assessment in oncologic imaging. The proposed framework has substantial potential to support precision oncology and facilitate the clinical translation of transparent AI-driven imaging biomarkers.
Keywords: Explainable Artificial Intelligence, Radiomics, Precision Oncology, Oncologic Imaging, Machine Learning, SHAP, Imaging Biomarkers, Prognostication, Survival Prediction, Artificial Intelligence.
View PDF Back