Robustness of Radiomic Features in [C]Choline and [F]FDG PET/CT Imaging of Nasopharyngeal Carcinoma: Impact of Segmentation and Discretization.

PURPOSE

Radiomic features are increasingly utilized to evaluate tumor heterogeneity in PET imaging and to enable enhanced prediction of therapy response and outcome. An important ingredient to success in translation of radiomic features to clinical reality is to quantify and ascertain their robustness. In the present work, we studied the impact of segmentation and discretization on 88 radiomic features in 2-deoxy-2-[F]fluoro-D-glucose ([F]FDG) and [C]methyl-choline ([C]choline) positron emission tomography/X-ray computed tomography (PET/CT) imaging of nasopharyngeal carcinoma.

PROCEDURES

Forty patients underwent [F]FDG PET/CT scans. Of these, nine patients were imaged on a different day utilizing [C]choline PET/CT. Tumors were delineated using reference manual segmentation by the consensus of three expert physicians, using 41, 50, and 70 % maximum standardized uptake value (SUV) threshold with background correction, Nestle's method, and watershed and region growing methods, and then discretized with fixed bin size (0.05, 0.1, 0.2, 0.5, and 1) in units of SUV. A total of 88 features, including 21 first-order intensity features, 10 shape features, and 57 second- and higher-order textural features, were extracted from the tumors. The robustness of the features was evaluated via the intraclass correlation coefficient (ICC) for seven kinds of segmentation methods (involving all 88 features) and five kinds of discretization bin size (involving the 57 second- and higher-order features).

RESULTS

Forty-four (50 %) and 55 (63 %) features depicted ICC ≥0.8 with respect to segmentation as obtained from [F]FDG and [C]choline, respectively. Thirteen (23 %) and 12 (21 %) features showed ICC ≥0.8 with respect to discretization as obtained from [F]FDG and [C]choline, respectively. Six features were obtained from both [F]FDG and [C]choline having ICC ≥0.8 for both segmentation and discretization, five of which were gray-level co-occurrence matrix (GLCM) features (SumEntropy, Entropy, DifEntropy, Homogeneity1, and Homogeneity2) and one of which was an neighborhood gray-tone different matrix (NGTDM) feature (Coarseness).

CONCLUSIONS

Discretization generated larger effects on features than segmentation in both tracers. Features extracted from [C]choline were more robust than [F]FDG for segmentation. Discretization had very similar effects on features extracted from both tracers.