Work

AI-assisted radiomics for classification of benign and non-benign right heart masses in 2D-echocardiography

Year: 2026

Type: article

Abstract: BACKGROUND: The rarity of right heart masses challenges diagnostic proficiency, while reproducibility is affected by the echocardiography operator. Artificial intelligence (AI)-based imaging tools may help address these limitations. METHODS: In this retrospective study (2013-2024), we enrolled surgical patients with right heart masses and obtained preoperative transthoracic (TTE) and transesophageal (TEE) echocardiographic images. Two-dimensional (2D) TTE (n = 98) and TEE (n = 87) images underwent radiomics analysis. Binary classification models were developed to differentiate benign from non-benign lesions using five machine-learning (ML) algorithms (decision trees, logistic regression, random forests, support vector machines (SVMs), extreme gradient boosting (XGBoost)). ML performance was compared with that of a deep-learning model based on the residual network (ResNet)-18 architecture using standard evaluation metrics such as the area under the curve (AUC). RESULTS: In 2D TTE analysis, ResNet-18 achieved the highest AUC (0.889), followed by XGBoost (0.836) and decision tree (0.815). ResNet-18 significantly outperformed SVM (P = 0.013) and logistic regression (P = 0.028), but showed no significant differences versus XGBoost (P = 0.408), decision tree (P = 0.429), or random forest (P = 0.053). In 2D TEE analysis, SVM achieved the highest AUC (0.959), followed by XGBoost (0.924) and random forest (0.906), with no significant differences among these models (all P > 0.05). ResNet-18 (AUC = 0.900) significantly outperformed only the decision tree (P = 0.027). CONCLUSION: ResNet-18 showed the highest TTE AUC and outperformed SVM and logistic regression, but was comparable to other ML models. SVM achieved the highest TEE AUC, with no significant differences among top models. These findings provide a preliminary AI benchmark for right heart mass diagnosis, though external validation is needed.

Source: Journal of Translational Medicine

DOI: 10.1186/s12967-026-08322-8