Program Information
Robustness of Radiomics On Head and Neck CTs: Intra- and Inter-Observer Contouring Effects
J Crosby*, T Miller , H Li , L Lan , D Ginat , M Giger , The University of Chicago, Chicago, IL
Presentations
SU-K-702-5 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: 702
Purpose: To compare quantitative radiomic texture features of head and neck tumors and their variations in predicting cancer progression across intra- and inter-observer contouring.
Methods: Our retrospective study was conducted under an approved IRB, HIPAA-compliant protocol. Computerized radiographic texture analysis was performed on the pre-treatment CT scans of 35 HPV-positive oropharyngeal squamous cell carcinomas as part of a response-adapted radiation volume de-escalation trial treated with induction chemotherapy. Initial contours throughout the tumor slices varied in number of pixels between slices. Second and third contours were independently drawn by a second observer using the initial contours for tumor location. Next, our Radiomics Texture Analysis (RTA) workstation computed texture features within each contoured region and ROC analysis was conducted yielding the corresponding AUC for each feature in the task of distinguishing between tumors that demonstrated progression and those that did not. The outputs were used in assessing the intra- and inter-observer comparisons. The third round of contouring was compared to the first round for inter-observer comparison, and to the second round for intra-observer comparison.
Results: For the inter-observer comparison, an average AUC percent change of 16.4% ±13.2% was found between the first and second round and 11.8% ± 8.6% between the first and third round. An average AUC percent change of 19.7% ± 13.5% resulted between the second and third contouring rounds for the intra-observer comparison.
Conclusion: Performance of the features in terms of assessing tumor progression was promising. While percent changes in the AUC values were greater than 10%, the normalized texture feature values were within error bars among the contouring rounds for both intensity- and spatial-based texture features. From these early results, it appears that multiple contouring rounds should be performed as verification given that the AUC change was larger for intra-observer than inter-observer comparisons (19.7± 13.5% vs. 16.4% ±13.2%).
Funding Support, Disclosures, and Conflict of Interest: Funded by NIH T32 EB002103, QIN U01CA195564, and an University of Chicago Radiomics grant. MLG is a stockholder in R2 technology/Hologic, receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi and Toshiba, and is a cofounder of and stockholder in Quantitative Insights. HL receives royalties from Hologic.
Contact Email: