Diagnostic value and interpretability of [

backgroundto assess the value of [18 F]FDG PET/CT radiomics in infective endocarditis (IE) diagnosis. We evaluated and collected [18 F]FDG PET/CT and clinical data of 447 patients, with suspected IE studied in 3 centers. Radiomic features were calculated and after dimensionality reduction, we performed: (1) univariate testing for assessing the discrimination power of clinical variables and radiomics; (2) a multivariate random forest-based model fed by radiomics to predict the outcome of PET/CT visual analysis; (3) a clustering-based radiomic model to predict final diagnosis; (4) a series of Logistic Regression (LR) models to assess the relative contribution of each criterion in relation with final diagnosis.

results9/17 clinical and 7/11 radiomics variables were able to univariately stratify patients. The random forest model accurately predicted PET/CT visual assessment in definite cases, providing a classification of doubtful cases resembling the “IE-negative” radiomic signature. The clustering-based analysis divided patients in two groups. LRs demonstrated that the richer the information fed into the model, the higher the performances: the models including radiomics performed better than the one solely including visual image assessment.

conclusionRadiomic signature, employing both supervised and unsupervised approaches, effectively described and differentiated [18 F]FDG PET/CT outcomes in a large IE cohort. The identification of specific signatures for equivocal PET/CT findings suggests that radiomic features can assist in interpreting ambiguous PET results, thus significantly impacting patient management. Clustering algorithm successfully associated patients with varying conditions, allowing for further assessment and characterization within the radiomics framework, potentially leading to risk score-based interpretations.