Textural features extracted from 18 FDG PET/CT images help discriminate prosthetic valve endocarditis from inflammatory disease

Introduction: Prosthetic valve endocarditis (PVE) is challenging to distinguish from non-specific inflammation (NSI) while this differential diagnosis is essential for patient management. We investigated whether radiomic features extracted from ¹⁸F-FDG PET/CT images of the prosthetic valve region could assist in the differential diagnosis.

Methods: PET/CT images were obtained 60 min after injection of 4 MBq/kg of ¹⁸F-FDG using a PET/CT scanner (GE Discovery 690), 3 min per bed position, from 29 patients with PVE and 88 patients with NSI. Images were reconstructed using 3D OSEM TOF including attenuation correction and analyzed using LIFEx (Nioche et al, Cancer Res 2018). For each patient, a region of interest including the valve was drawn based on the CT scan by combining thresholding and dilation operations. The region was copied to the PET scan and 42 features (SUVmax, histogram based features and textural features) were extracted. Univariate ROC analyses were performed for each feature using 2/3 of the whole data set as a training set (TS) to determine the optimal cut-off defined as the one maximizing the Youden index (sensitivity+specificity-1) to separate PVE from NSI. The Youden index for that cut-off was then calculated on the remaining 1/3 data set, called external set (ES). This was performed 20 times with a different random split of the whole data set. In addition, linear models and scores to identify PVE from NSI using 2 features were systematically screened on the TS and assessed on the ES. The process was also repeated 20 times with different random splits.

Results: Using ROC analysis, SUVmax yielded the highest performance to identify PVE from NSI, with an area under the curve (AUC) of 0.83±0.03 on the ES. Using an SUVmax cut-off of 4.5±0.3, the sensitivity±1SD/specificity±1SD for PVE detection calculated on ES were 73±8%/76±17%. When combining 2 features, the most promising model was a score combining High Grey-level Zone Emphasis (HGZE) with SUVmax. Patients with at least one “risk” factors (SUVmax > 4.5±0.3 and/or HGZE > 340±25) corresponded to 85±12% of the patients with PVE in the ES, which was significantly higher (p<0.05) than the sensitivity obtained when using SUVmax only. Patients with no risk factor (SUVmax &#8804;4.5±0.3 and HGZE &#8804; 340±25) corresponded to 65±12% of the patients with NSI, which was not significantly different from the specificity obtained when using SUVmax only (p>0.05). Yet, SUVmax and HGZE were highly correlated (Pearson r of 0.94). This suggests that their combination mostly increases the reliability of the assessment of the maximum metabolic activity within the lesion instead of actually reflecting the spatial distribution of the metabolic activity on top of its magnitude.

Conclusions: These results suggest that accounting for one or more radiomic features might be useful to identify patients with PVE. Increasing the size of the cohort will make it possible to investigate whether machine learning approaches involving these features and more sophisticated ones can yield better classification performance.