18F]FDG-PET/CT radiomics in patients suspected of infective endocarditis

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. AIM [18F]FDG-PET/CT is part of the diagnostic algorithm for IE diagnosis. Increased [18F]FDG uptake with focal and heterogenous pattern at valve, intravalvular or perivalvular at visual analysis is consistent with IE. Diffuse, homogeneous or low valvular [18F]FDG uptake make diagnosis more challenging. Semiquantitative parameters may be of value in such case of equivocal PET findings; however, they are still not validated in IE. In this study we aim to assess the value of [18F]FDG PET/CT radiomics in IE diagnosis. Further, we build a model for radiomics-based prediction of PET/CT findings, patient classification and stratification as well as prediction of the final diagnosis. Materials and Methods We evaluated a series of [18F]FDG PET/CT scans in 447 patients (M:F =284:163, mean age 67± 16yrs), with suspected IE (519 valves, NVE = 109, PVE = 410), studied in 3 different centers between January 2015- 2020. Clinical, surgical data, antimicrobial treatment, microbiology and biochemistry, imaging and the DUKE/2015 ESC classification were collected. PET/CT images were semiautomatically segmented (Advantage Workstation, GE) and texture features extracted by LIFEx software. For the analysis we used absolute correlation exclusion criteria and PCA based dimensionality reduction, MANOVA test and LR for multivariate testing. Prior to model building by Random Forest (80% training sets, 20% test), we applied covariance matrix for correlated feature removal and SMOTE for preprocessing the imbalanced dataset. Results MANOVA and LR showed a positive contribution of radiomics in predicting PET/CT results and IE diagnosis, with a different signature in IE-positive/IE-negative patients (80% in training, 70% in validation). Of interest, the signature of patients with equivocal PET/CT findings was similar to IE-negative signature. Clustering-based stratification identify in two groups, one with milder disease presenting weak or no [18F]FDG uptake and one with more severe disease. Our LR models with incremental complexity (Table 1 and 2) demonstrated that the richer the information fed into the model the higher the performances, reaching 90% of AUC. However, the performance of model M5 and M6 is almost equal, suggesting a limited contribution of radiomics in classifying IE. Conclusion [18F]FDG PET/CT radiomics provide a limited, yet positive, contribution in the classification of EI. Nevertheless, radiomics was fundamental in defining PET outcome, thus it could support visual imaging assessment in particular when equivocal [18F]FDG findings are present. Further steps focusing on refinement of the IE diagnostic criteria, on explainable analysis on positive/negative patients to be transferred in equivocal cases. Ultimately, the identification of radiomic signature would help to define thresholds to discriminate between mild infection and severe IE, in a risk score fashion. Abstract Table 1 Abstract Table 2