ETD

• artificial intelligence
• cardiac masses
• cardiac thrombus
• cardiac tumors
• deep learning
• machine learning
• myxoma
• vegetation

Background: Intracardiac masses comprise a heterogeneous group of diseases with challenging diagnostic characterization. Differentiating tumors from thrombi is particularly relevant for appropriate management strategies. Cardiac ultrasound modalities, particularly transesophageal echocardiography (TEE), are the first-line diagnostic tool, but more definitive diagnosis requires advanced imaging modalities or histology. Artificial intelligence (AI), particularly machine learning and deep learning algorithms, are emerging as a transformative tool in echocardiography, enhancing image interpretation. Here we implemented and evaluated an AI model for characterizing cardiac masses from TEE images, comparing its performance with magnetic resonance imaging or post-surgical histology as gold standards.
Methods: A total of 53 series of TEE images of intra-atrial masses obtained from 43 patients were analyzed. Twenty-one masses were ultimately diagnosed as cardiac tumors (76% myxomas), while the remaining were classified as thrombi, or endocarditis vegetations. Regions of interest were selected from single frames by two experienced echocardiographers blinded to the final diagnosis. A repeated nested cross-validation (CV) scheme was used, consisting of 100 repetitions of a 5-fold CV with stratification at the patient level based on target labels. Radiomic features were z-score-normalized at the training set level to avoid performance bias. Hyperparameter selection was done using a 5-repetitions 2-fold CV. The model included a feature selection step, followed by regularized logistic regression. To assess whether results were obtained by chance, a permutation test was conducted by shuffling the target labels in each external CV repetition and comparing the area under the receiver operating characteristic curve (AUC-ROC).
Results: The AUC-ROC for the standard dataset was (median [Interquartile range]) 0.64 [0.60, 0.67], compared to 0.51 [0.45, 0.58] for the permuted model (p=3.75 × 10⁻²⁰, Mann–Whitney U test). The model performance metrics included a true positive rate of 0.54 [0.51, 0.57], a true negative rate of 0.64 [0.62, 0.69], a positive predictive value of 0.50 [0.46, 0.53], a negative predictive value of 0.70 [0.68, 0.72], and 0.61 [0.59, 0.63] accuracy.
Conclusions: AI integration in echocardiography imaging can significantly enhance diagnostic precision in distinguishing cardiac tumors from other cardiac space-occupying lesions, thus contributing to improved clinical decision-making and to an appropriate second-tier resource utilization.