Extracting and Selecting Robust Radiomic Features from PET/MR Images in Nasopharyngeal Carcinoma

This work aims to study the variation, robustness, and feature redundancy of PET/MR radiomic features in the primary tumor of nasopharyngeal carcinoma (NPC).

PET/MR scans of 21 NPC patients were used in this study. The primary tumor volumes were defined using PET, T2-weighted-MR (T2-MR), and diffusion-weighted MR (DW-MR) images. A random-dilation-erosion method was used to simulate 10 sets of tumor volumes for identifying features invariant with manual segmentation uncertainties. Feature robustness was evaluated against imaging modalities, pixel sizes, slice thickness, and grey-level bin sizes using intraclass correlation coefficient (ICC) and spearman correlation coefficient. Feature redundancy was analyzed using the hierarchical cluster analysis.

Voxel size of 0.5 × 0.5 × 1.0 mm³ was found optimal for robust feature extraction from PET and MR. Normalized grey level of 64 and 128 was suggested for PET and MR, respectively. The features from wavelet-transformed images were less stable than those from the original images. The robustness analysis and volume correlation analysis identified 335 (62.04 %) PET features, 240 (44.44 %) T2-MR features, and 366 (67.78 %) DW-MR features. The cluster analysis grouped PET, T2-MR, and DW-MR features into 106, 83, and 133 representative features, respectively.

The present study analyzed and identified robust features extracted from tumor volumes on PET/MR, which can provide guidance and promote standardization for PET/MR radiomic studies in NPC.