CMR-based radiomics to predict cardiac masses malignancy

Technical developments in medical imaging have significantly improved the diagnosis of cardiac masses (CMs). While histological examination remains the gold standard, multimodal imaging has become crucial in clinical practice. Compared to other techniques, CMR offers superior tissue characterization, though current methods are largely qualitative. Radiomics enables high-dimensional quantitative image analysis and has demonstrated promise in oncological imaging. However, its application in cardiac imaging remains largely unexplored.

This study aimed to evaluate the potential of radiomics-based analysis using CMR cine (CINE) and late gadolinium enhancement (LGE) sequences for differentiating benign from malignant cardiac masses.

A retrospective study was conducted on patients from three hospitals who underwent CMR for suspected CMs. Final diagnoses were confirmed through histological examination or radiological resolution following anticoagulation in thrombus cases. Radiomics features were extracted from CINE and LGE sequences using the PyRadiomics library, after preprocessing steps such as noise reduction, intensity correction, and standardization. Three models were developed: one using CINE data, one using LGE data, and one combining both. Stability analysis was performed to refine feature selection, followed by the evaluation of multiple classification models. Radiomics performance was assessed using a training and independent test dataset.

The study included 170 patients, with 52% diagnosed with benign masses and 48% with malignant masses. Stability analysis reduced features from 851 to 240 (CINE) and from 474 to 169 (LGE). The best-performing classifier varied across models: AdaBoost for CINE and LGE models, and k-NN for the combined model. On an independent test set, the models achieved 84.62% accuracy for CINE, 81.82% for LGE, and 90% for the combined model.

Radiomics analysis of CMR cine and post-contrast imaging can effectively differentiate between benign and malignant cardiac masses with high accuracy. This approach offers a promising complementary tool for non-invasive cardiac mass evaluation, particularly in settings where expertise in CMR interpretation is limited. However, further validation with larger datasets is required to enhance clinical applicability.