Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Investigating Disease-Stage Dependence of MR-Based Radiomic Feature Reproducibility Against Image Perturbations in Nasopharyngeal Carcinoma

T Cheung, S Lam, J Zhang*, J Cai, Hong Kong Polytechnic University, Hong Kong, CN


PO-GePV-M-64 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To investigate disease-stage dependence of MR-based radiomic features reproducibility against perturbations in nasopharyngeal carcinoma (NPC) patients.

Methods: A total of 142 NPC patients were retrospectively grouped into stage-III (n=96) and stage-IV (n=46) cohorts. Forty combinations of image perturbations, including translation, rotation, noise addition, and segmentation randomization, were applied on the contrast-enhanced T1-weighted (CET1w) and T2-weighted (T2w) MR images of each patient to simulate variabilities in patient scanning position, imaging noise and tumor segmentation. 107 radiomic features were calculated from gross-tumor-volume per perturbation combination, including raw, Laplacian-of-Gaussian (LoG) (sigma=1,3,5 mm) and wavelet features. Intra-class correlation coefficient (ICC) was used to assess reproducibility of each individual feature in the two cohorts separately. Median ICC (mICC) was then calculated for each feature subclass. Ratio of mICC of each feature subclass for stage-IV to that for stage-III was used to assess direction and magnitude of stage-dependence; Mann-Whitney U test was adopted to examine presence of statistically significant difference of mICC between cohorts. Besides, proportion of stage-dependent robust features (ICC>0.9) was calculated.

Results: In general, stage-IV presented lower mICC values than stage-III for raw, LoG (1-mm) and wavelet features, but higher for LoG features with larger kernel size (3-mm and 5-mm). Particularly, CET1w wavelet features exhibited 40-60% lower mICC in stage-IV compared to stage-III (p-value: 0.0001-0.005). Further, among the robust wavelet radiomic features (ICC>0.9), 90% expressed stage-dependence. By contrast, T2w features generally displayed weaker stage-dependence characteristics than T1w features.

Conclusion: This study demonstrated that types of radiomic features, the use and setting of filters, played an influential role in stage-dependence of radiomic feature reproducibility. Particularly, CET1w wavelet feature reproducibility exhibited significantly strong stage-dependence in NPC tumor, potentially ascribed to the intra-tumoral heterogeneity between stages. Association of stage-dependence and tumor volume will be evaluated in the next phase of the study.

Funding Support, Disclosures, and Conflict of Interest: This work is funded by Innovation and Technology Fund (ITS/080/19), the Innovation and Technology Commission, Project of Strategic Importance (P0035421), The Hong Kong Polytechnic University, and Shenzhen-Hong Kong-Macau S&T Program (Category C) (SGDX20201103095002019), Shenzhen Basic Research Program (R2021A067)


Segmentation, MRI, Statistical Analysis


IM- MRI : Radiomics

Contact Email