An exploration of ultrasound-based radiomics combined with machine-learning in the diagnosis of fetal ventricular disproportion caused by cardiovascular malformations

Determining whether ventricular disproportion is physiological or caused by structural malformations during the fetal period can often be challenging. Radiomics has shown potential in cardiac structure and function evaluation. This study aimed to explore the value of fetal echocardiography-based radiomics in the evaluation of fetal ventricular disproportion. In this single-centre retrospective diagnostic study, 82 fetuses with ventricular disproportion and 120 fetuses with normal ventricular size were included. And 82 fetuses with ventricular disproportion were followed up for outcomes and diagnoses. 294 radiomic features of myocardium were extracted from 202 still end-diastolic 4-chamber view of echocardiographic images. Radiomics-based machine learning models were created to distinguish controls from fetuses with ventricular disproportion, and to differentiate ventricular disproportion with from without cardiovascular malformations. Of 294 extracted radiomic features, 27 met the reproducibility threshold. For the one-vs-all multiclass classifiers, the best-performing models achieved AUCs of 0.652, 0.701, and 0.609 for Control vs. Rest, Group A vs. Rest, and Group B vs. Rest, respectively. For the binary classification of Group A versus Group B, the combined model achieved the highest AUC of 0.742 (95% CI 0.633–0.852), with no clear gain over clinical parameters alone. Radiomics combined with machine learning demonstrate preliminary feasibility in distinguishing fetal ventricular disproportion with and without cardiovascular malformations. Further external validation in larger multicentre prospective cohorts is required. Not applicable.