Prediction of HPV Status in the Head and Neck Cancer Patient Using CT Derived Radiomic Features: The Impact of CT Scanner
R Reiazi1,2*, C Arrowsmith1, M Welch3, F Abbas-aghababazadeh1,C Eles1, T Tadic1,2,3, A Hope1,2,3, S Bratman1,2,3, B Haibe-kains1,2,3,4,4,5,6,(1) Radiation Medicine Program, Princess Margaret Cancer Centre, University Health Network, (2) Department of Medical Biophysics, University of Toronto, (3) Department of Radiation Oncology, University of Toronto, (4) Ontario Institute for Cancer Research, (5) Department of Computer Science, University of Toronto, (6) Vector Institute
Presentations
PO-GePV-M-45 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: In this study, we assessed the impact of scanner choice on the computed tomography (CT)-derived radiomic features to predict association of oropharyngeal cancer (OPC) with human papillomavirus (HPV).
Methods: A total of 1,294 OPC patients with known HPV status were retrospectively included in this study. In this patient cohort, CT images were acquired with two different scanner types. We extract a total of 1,874 IBSI (International Biomarker Standard Initiation) compliant radiomic features from the primary gross tumor volume (pGTV) from each patient. Feature selection has been done by Maximum Relevance Minimum Redundancy Ensemble (mRMRe)3 technique. The dataset was subsequently stratified by CT scanner type (Toshiba scanner, GE scanner, and Mix) and splitted into train (adjusted for class imbalance) and test sets with the proportion of 75/25. Next, a random forest classifier was trained on the following configurations: (1)Toshiba-Toshiba, (2)GE-GE, (3)Toshiba-GE, (4)GE-Toshiba, (5)Mix-Mix, (6)Toshiba-Mix, (7)GE-Mix, (8)Mix-Toshiba and (9)Mix-GE.
Results: The highest and lowest mean AUC (area under the curve) values were 0.79 (p-value<0.001) and 0.70 (p-value<0.001) and obtained with Toshiba-Mix and Toshiba-GE respectively (Figure 1). For models trained on one scanner type, the highest and lowest result obtained when they were tested on Mix sample (i.e GE-Mix [0.75,p-value<0.001], Toshiba-Mix [0.79, p-value<0.001]) and other scanner types (i.e GE-Toshiba [0.73,p-value < 0.001], Toshiba-GE [0.70,p-value < 0.001]) respectively. We identified that there is bias in the results in favor of one scanner type (Toshiba). Models either trained or tested on the Toshiba scanner resulted in higher predictive value compared to other scanners in all the same configurations.
Conclusion: In this study, scanner type affects the prediction accuracy of the HPV status using hand-engineered radiomics features. The optimal prediction accuracy was achieved when the training set included only one specific type of scanner which reflects a bias in radiomic features towards the scanner type.
Funding Support, Disclosures, and Conflict of Interest: Research reported in this study was supported by Canadian Institute of Health Research (CIHR) under grant number: 426366
Keywords
Not Applicable / None Entered.
Taxonomy
IM- Dataset Analysis/Biomathematics: Machine learning
Contact Email