Radiomic Analysis of Positron-Emission Tomography and Computed Tomography Images to Differentiate between Multiple Myeloma and Skeletal Metastases

CONTEXT: Multiple myeloma and extensive lytic skeletal metastases may appear similar on positron-emission tomography and computed tomography (PET-CT) in the absence of an obvious primary site or occult malignancy. Radiomic analysis extracts a large number of quantitative features from medical images with the potential to uncover disease characteristics below the human visual threshold. AIM: This study aimed to evaluate the diagnostic capability of PET and CT radiomic features to differentiate skeletal metastases from multiple myeloma. SETTINGS AND DESIGN: Forty patients (20 histopathologically proven cases of multiple myeloma and 20 cases of a variety of bone metastases) underwent staging 18F-fluorodeoxyglucose PET-CT at our institute. METHODOLOGY: A total of 138 PET and 138 CT radiomic features were extracted by manual semi-automatic segmentation and standardized. The original dataset was subject separately to receiver operating curve analysis and correlation matrix filtering. The former showed 16 CT and 19 PET parameters to be significantly related to the outcome at 5%, whereas the latter resulted in 16 CT and 14 PET features. Feature selection was done with 7 evaluators with stratified 10-fold cross-validation. The selected features of each evaluator were subject to 14 machine-learning algorithms. In view of small sample size, two approaches for model performance were adopted: The first using 10-fold stratified cross-validation and the second using independent random training and test samples (26:14). In both approaches, the highest area under the curve (AUC) values were selected for 5 CT and 5 PET features. These 10 features were combined and the same process was repeated. STATISTICAL ANALYSIS USED: The quality of the performance of the models was assessed by MSE, RMSE, kappa statistic, AUC, area under the precision-recall curve, F-measure, and Matthews correlation coefficient. RESULTS: In the first approach, the highest AUC = 0.945 was seen with 5 CT parameters. In the second approach, the highest AUC = 0.9538 was seen with 4 CT and one PET parameter. CT neighborhood gray-level different matrix coarseness and CT gray-level run-length matrix LGRE were common parameters in both approaches. Comparison of AUC of the above models showed no significant difference (P = 0.9845). Feature selection by principal components analysis and feature classification by the multilayer perceptron machine-learning model using independent training and test samples yielded the overall highest AUC. CONCLUSIONS: Machine-learning models using CT parameters were found to differentiate bone metastases from multiple myeloma better than models using PET parameters. Combined models using PET and CECT data showed better overall performance than models using only either PET or CECT data. Machine-learning models using independent training and test sets were performed on par with those using 10-fold stratified cross-validation with the former incorporating slightly more PET features. Certain first- and second-order CT and PET texture features contributed in differentiating these two conditions. Our findings suggested that, in general, metastases were finer in CT and PET texture and myelomas were more compact.