(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
To investigate the utility of network-based analysis in identifying stable and reproducible features and validating predictive models.
Methods:
A set of 132 radiomic features were extracted on contrast enhanced computed tomography (CT) scans, reconstructed through two different methods, using the CERR radiomics toolbox. The two radiomic feature sets were computed from 47 lung nodules. For each feature set, a regularized partial correlation network was constructed, resulting in a form of sparsity. The commonality of the resultant two networks was assessed. To apply this idea to outcomes modeling, we developed a novel K-means algorithm coupled with the optimal mass transport theory, which was tested on CT radiomic features extracted from tumor regions of 77 head and neck cancer patients that were downloaded from The Cancer Imaging Archive (TCIA).
Results:
Two radiomic networks were constructed using two different kernels. The largest connected component from each network was compared; both consisted of 15 radiomic features and 10 features were common, showing the strong reproducibility of these radiomic features between different reconstruction methods. To further test the reproducibility of these features on phantom data, the two largest components were merged, making a larger network with 20 unique radiomic features. A simulation test using the Wasserstein distance computed on the merged network showed the stability of the 20 radiomic features. The partial correlation analysis on TCIA head and neck cancer data resulted in 3 connected components. The Wasserstein K-means clustering algorithm led to two clearly separable clusters. Significant differences in tumor subsite and HPV status were found between the two radiomic clusters with extended Fisher’s exact test p=0.0063 and p=0.0012, respectively.
Conclusion:
We showed that a network-based analysis enables identifying reproducible radiomic features. This was validated using phantom data and external data via the Wasserstein distance metric and the proposed Wasserstein K-means clustering method.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
