An assessment of PET and CMR radiomic features for detection of cardiac sarcoidosis

Abstract

Background: Visual interpretation of PET and CMR may fail to identify cardiac sarcoidosis (CS) with high specificity. This study aimed to evaluate the role of [18F]FDG PET and late gadolinium enhancement (LGE)-CMR radiomic features in differentiating CS from another cause of myocardial inflammation, in this case patients with cardiac-related clinical symptoms following COVID-19. Methods: [18F]FDG PET and LGE-CMR were treated separately in this work. There were thirty-five post-COVID-19 (PC) and forty CS datasets. Regions of interest were delineated manually around the entire left ventricle for PET and LGE-CMR datasets. Radiomic features were then extracted. The ability of individual features to correctly identify image data as CS or PC was tested to predict clinical classification of CS vs. PC using Mann–Whitney U-tests and logistic regression. Features were retained if P-value <0.00053, AUC >0.5 and accuracy >0.7. After applying correlation test, uncorrelated features were used as a signature (joint features) to train machine learning classifiers. For LGE-CMR analysis, to further improve the results, different classifiers were used for individual features besides logistic regression and the results of individual features of each classifier were screened to create a signature that include all features that followed the previously mentioned criteria and use them as input for machine learning classifiers. Results: The Mann–Whitney U-tests and logistic regression were trained on individual features to build a collection of features. For [18F]FDG PET analysis, the maximum target-to-background ratio (TBRmax) showed high area under the curve (AUC) and accuracy with small P-values (<0.00053) but the signature performed better (AUC 0.98 and accuracy 0.91). For LGE-CMR analysis, Gray Level Dependence Matrix (gldm)-Dependence Non-Uniformity showed good results with small error bars (accuracy 0.75 and AUC 0.87). However, by applying a Support Vector Machine classifier on individual LGE-CMR features and creating a signature, a Random Forest classifier displayed better AUC and accuracy (0.91 and 0.84, respectively). Conclusion: Using radiomic features may prove useful in identifying individuals with CS. Some features showed promising results to differentiate between PC and CS. By automating the analysis, the patient management process can be accelerated and improved.