A multi-organ cancer study of the classification performance using 2D and 3D image features in radiomics analysis.

The purpose of this study was to investigate the predictive performance of 2D and 3D image features across multi-organ cancers using multi-modality images in radiomics studies. In this retrospective study, we included 619 patients with three different cancer types (intrahepatic cholangiocarcinoma (ICC), high-grade osteosarcoma (HOS), pancreatic neuroendocrine tumors (pNETs)) and four clinical end points (early recurrence (ER), lymph node metastasis (LNM), 5-year survival and histologic grade). The image features included fifty-eight 2D image features and fifty-eight 3D image features. The 3D image features were extracted based on the 3D tumor volumes. The 2D image features were extracted based on 2D tumor region, which was the layer with the maximum tumor diameter within the 3D tumor volume. The predictive performance of individual 2D and 3D image feature was measured using the area under the receiver operating characteristic curve (AUC) with univariate analysis. Radiomics signatures were further developed using multivariable analysis with 4-fold cross-validation method. Using univariate analysis, we found that more 3D image features showed the statistically predictive capabilities than 2D image features across all the included cancer types. By comparing the predictive performance of radiomics signatures developed by 2D and 3D image features, we observed better prediction performance in radiomics signatures based on 3D image features than those based on 2D image features for patients with ICC and HGO. Meanwhile, the signatures based on 2D and 3D image features performed closely in the pNETs dataset with the clinical end point of the histologic grade. The reason for this inconsistent result might be that the gross tumor volumes of pNETs were generally small. The tumor heterogeneity was mostly presented in the middle several layers within the tumor volume. Both 2D and 3D image features have certain predictive capacities. By contrast, the 3D image features show better or close predictive performance than 2D image features using both univariate analysis and multivariate analysis. In brief, 3D image features are recommended in radiomics studies.