Program Information
Robustness Against Artifacts, Reliability and Reproducibility of CT Radiomics Features in Head and Neck Cancer
M Crispin Ortuzar*, J O Deasy, A Apte, Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
WE-F-205-9 (Wednesday, August 2, 2017) 1:45 PM - 3:45 PM Room: 205
Purpose: The sensitivity of radiomics features to factors such as image resolution, time difference, or image artifacts can compromise the predictive power of the signatures. In this study we assess the sensitivity of head-and-neck CT radiomics features to such effects using a large clinical dataset.
Methods: 100 head and neck patients who received chemoradiotherapy treatment were analyzed. All patients received pre-treatment FDG and FMISO PET/CT on different days. Two copies of the FDG PET/CT scans were available, one from the treatment planning system (CT₁) and one from the PACS system (CT₂, coarser resolution). A total of 123 lesions delineated by physicians and with a volume larger than 10 cc were analyzed. Slices with artifacts were removed, and over 200 radiomics features were calculated. Intraclass correlation coefficients (ICC) were used to assess the reliability (ICCrel+/-, CT₁ vs CT₂ with and without reslicing), reproducibility (ICCrep, CT₂ vs FMISO CT, using only scans with the same resolution) and robustness against artifacts (ICCa, CT₁ with vs without artifacts).
Results: The mean reliability was ICCrel+=0.58 with reslicing, and ICCrel-=0.75 without reslicing. The mean reproducibility was ICCrep=0.55, while the mean robustness against artifacts was ICCa=0.80. The four metrics were higher in 3D features compared to 2D features. Reslicing was observed to significantly improve the reliability in some cases (e.g. RLM RLN) and significantly worsen it in others (e.g. RLM LRE). The only feature with all four metrics above 0.9 was the total energy of the patch-wise Haralick entropy texture map.
Conclusion: CT artifacts, reslicing techniques and changes between repeat CTs on different days for the same patient may lead to significantly different radiomic metrics in head and neck cancer. Variability differed based on the individual radiomic feature in question, and therefore should be taken into account in building predictive models.
Contact Email: