Beschreibung:
Abstract Introduction FDG PET/CT allows for a better sensitivity in the prosthetic valve endocartitis (PVE) diagnostic when integrated to ESC 2015 criteria, but visual image analysis results in a weaker specificity and is subject to an inter-observer variability. We therefore aimed to evaluate the interest of quantitative analysis using radiomics and machine learning of FDG PET/CT scans in the PVE diagnostic. Material and methods Between 2015 and 2021, patients referred for a FDG PET/CT in our nuclear medicine department with suspected PVE were retrospectively included. The initial development of the model was focused on aortic prosthetic valve (aPV). The aPV was segmented and 31 radiomics features were extracted using the IBSI compliant PyRadiomics framework. Radiomics features were first tested by shuffling 50 times the signal within the aortic segmentation and non-contributive (i.e. identical results within 2×1,96σ over 50 iterations) were excluded. Correlated features were further removed using the variable inflation factor blinded to outcome and remaining features were standardized. Four machine learning algorithms (Ridge and LASSO logistic regression, support vector classifier and random forest) were evaluated and tuned through the use of a training database of patients with aPV included from 2015 to 2019 (excluding positive patient with a mitral and aortic valve). The procedure was further tested through 100 loops on an additional cohort of patients with only aPV included after 2019. ROC curves were subsequently computed and sensitivity was derived based on a fixed specificity of 0.7. Gold standard consisted in an expert consensus from the Endocarditis team. Primary objective was to assess the diagnostic performances of our combined approach using radiomics features and clinical features related to the PET exam (i.e time between aPV implantation and FDG PET/CT, time between antibiotics initiation and FDG PET/CT, extracardiac positive foci, spleen uptake and bone marrow uptake greater than liver uptake). Results 108 patients were included, for a total of 65 definite PVE and 43 rejected PVE according to the expert consensus. The four algorithms were trained on a total of 68 patients and further tested on a cohort of 40 patients. The performance metrics are reported in the table. Support vector classifier achieved the best scores with an AUC of 0.79±0.01 (sensitivity 0.74±0.03; specificity 0.7). When adding clinical features, AUC was 0.82±0.02 (sensitivity 0.78±0.02; specificity 0.7). Conclusion When analyzed with our machine learning-based algorithms, FDG PET/CT reached acceptable diagnostic performances in terms of sensitivity for a specificity corresponding to the results reported by the ENDOPET study (1). These preliminary results obtained on a small test dataset suggest that an artificial intelligence-based algorithm may then guide the final diagnosis especially in this area of subjective visual assessment of PVE. Funding Acknowledgement Type of funding sources: None.