"AI-Driven Malignancy Detection in Cardiac Tumors via T1-Weighted MRI Imaging"

Meri Ferretti1, Michele Pagliaccia2, Andrea Baggiano3, Gianluca Pontone4, Francesco Angeli5, Matteo Armillotta5, Luca Bergamaschi5, Luigi Lovato6, Carmine Pizzi5, Valentina Corino1
1Politecnico di Milano, 2Cardiology and Cardiovascular Pathophysiology, University of Perugia and Hospital S. Maria della Misericordia, Perugia, Italy, 3Department of Clinical Sciences and Community Health (E. Gherbesi, S.C., A.B., G.P.), University of Milan, Italy, 4IRCCS Centro Cardiologico Monzino, 5Department of Medical and Surgical Sciences, Alma Mater Studiorum, University of Bologna, 40138 Bologna, Italy, 6Department of Pediatric and Adult Cardio-Thoracovascular, Oncohematologic and Emergencies Radiology Unit, IRCCS Azienda Ospedaliero-Universitaria di Bologna, Bologna, Italy


Abstract

Aim: We propose a novel tool for early differentiation between primary and metastatic malignant cardiac tumors, a rarely encountered pathology, with primary tumors accounting for only 0.002%–0.3% of cases among all biopsy series. Leveraging contrast-free T1-weighted cardiac cine MRI, this tool is designed for reproducibility and easy clinical integration, reducing reliance on advanced imaging and expert interpretation while also improving accessibility for patients with contraindications to gadoliniumbased agents. Methods: We analyzed 36 tumors, extracting volumetric radiomic features from segmented cardiac MRI. To identify the most informative features, we computed a correlation analysis and then evaluated three additional feature selection methods: Mann-Whitney significance testing, minimum Redundancy-Maximum Relevance, and Least Absolute Shrinkage and Selection Operator (LASSO) regression. The selected features were then fed to several machine learning algorithms, with performance assessed through 10-fold cross-validation. Model explainability was evaluated using permutation importance and Shapley Additive explanations (SHAP). Results: The combination of correlation analysis, LASSO-based feature selection, and a Support Vector Machine classifier yielded superior performance, with a mean validation accuracy of 0.90±0.15 over 10 different folds. After retraining the model on the full training dataset and evaluating it on an external cohort of six patients, it achieved an overall accuracy of 0.83 and demonstrated perfect sensitivity (100%) with only one false positive. Seven radiomic features predictive of malignancy have been utilized in the final model. Lesion flatness resulted the most influential feature (importance coefficient = 0.17); specifically, higher lesion flatness was linked to a lower likelihood of metastasis, suggesting that metastatic tumors may exhibit more irregular shapes, potentially due to rapid and disorganized growth. Conclusion: These findings support the potential of our AI-based tool to accurately classify primary versus metastatic cardiac tumors using standard MRI, without the need for contrast agents and specialized expertise.