Click here to

Session: Data Science, Radiomics, and Computing [Return to Session]

Spatial Reconstruction of Statistically Significant Radiomics Signatures Using 3D Wavelet Decomposition in Tumors of Oropharyngeal Cancer

H Bagher-Ebadian*, F Siddiqui, A Ghanem, S Zhu, M Lu, B Movsas, I Chetty, Henry Ford Health System, Detroit, MI

Presentations

SU-E-TRACK 6-7 (Sunday, 7/25/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: We applied 3D-wavelet-decomposition to spatially reconstruct statistically significant radiomics-features within images of the tumor for patients with oropharyngeal-squamous-cell-carcinomas(OPSCC’s), for characterization of HPV.

Methods: One-hundred-twenty-eight OPSCC-patients with known HPV-status (60-HPV+ and 68-HPV-, confirmed by immunohistochemistry-P16-protein-testing) were retrospectively studied. Three-dimensional-wavelet-decomposition analysis was performed on the contrast-enhanced CT-images of patient gross-tumor-volumes to decompose the tumor’s intensity and spatial-frequency information into 3D-decomposition levels with series of orthonormal high-pass and low-pass wavelet-coefficients (WCs). Log-Energy-Entropy of the WCs for all the decomposition levels were calculated as radiomic-features. A Least-Absolute-Shrinkage-and-Selection-Operation (Lasso) technique combined with a Generalized-Linear-Model (Lasso-GLM) was applied on these features to identify and rank the optimal feature subsets with most representative information for prediction of HPV-status. Lasso-GLM classifier was constructed and validated using random-permutation-sampling of the features. Average Area-Under-Receiver-Operating-Characteristic (AUC), and Positive and Negative Predictive-values (PPV/NPV) were computed to estimate the generalization-error and prediction performance of the classifier. Ultimately, the optimal radiomic-features selected by the Lasso-analysis were used to select the significant wavelet-sub-bands and their respective WCs to perform inverse-wavelet-transform to reconstruct the tumor zones containing the highest information towards prediction of HPV-status.

Results: Four entropy-based features from two decomposition-levels, including two low-frequency (LLL3/LHL3) and two high-frequency components (HHL1/HLL1) were found to be statistically significant discriminators between HPV-negative and HPV-positive cohorts respectively. The wavelet-based-Lasso-GLM classifier’s prediction performance was: AUC/PPV/NPV=0.798/0.745/0.826. Results imply that tumor zones with smaller-components (higher-details/spatial-frequencies: components-sizes<~2mm) contain important radiomic information for characterization of HPV+ patients while tumor zones with larger-components (low-spatial-frequencies: components-sizes>~8mm) contain information for classification of HPV-cases. This study also confirms that compared to the central-zones of tumors, peritumoral-regions contain more information for characterization of the HPV-status.

Conclusion: Results suggest that radiomics can be used to discriminate spatial-regions of importance toward the contribution of HPV. By associating this information with tumor pathology, one can potentially link radiomics to underlying biological mechanisms.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by a grant from Varian Medical Systems (Palo Alto, CA)

Handouts

    Keywords

    Adaptive Sampling, Feature Extraction, Image Processing

    Taxonomy

    IM/TH- Informatics: Informatics in Imaging (general)

    Contact Email

    Share: