(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: To quantitatively compare the stability of radiomic features extracted from 2-D and 3-D MR images of brain metastases to inform data acquisition for reliable and reproducible radiomics studies.
Methods: Under IRB approval, a retrospective cohort of 29 patients with brain metastases who had contrast-enhanced T1-weighted MR images acquired using 2-D spin echo (SE) and 3-D spoiled gradient echo (SPGR) sequences within one exam included in this analysis. Tumor volumes were contoured using semi-automated methods by experienced physicians. 2-D radiomic features (0.4297×0.4297×5-mm spatial normalization (SN), 64-bin intensity discretization (ID)) and 3-D radiomic features (3×3×3-mm SN, 64-bin ID) were extracted using PyRadiomics. Coefficient of variation (CV) was computed for all 2-D and 3-D features extracted from both the 2-D SE and 3-D SPGR MR images. Though not representative of MRI, CV was computed for features from synthetic white noise images with resolution matching patient images as proof of concept.
Results: Of the 100 2-D and 3-D radiomic features selected a priori, using a robustness threshold of CV<10%, 11 were robust for both 2-D and 3-D image sets; 4 were robust only for 2-D image sets and 2 were robust for only 3-D image sets. All features that met the robustness threshold in at least one image set had CV<13% for the other. The CV was smaller for 79% of 2-D features (median CV difference 4.86%) and 72% of 3-D features (median CV difference 4.40%) when derived from 2-D SE image sets vs. 3-D SPGR. Synthetic white noise images yielded similar results.
Conclusion: This work indicates that radiomic features of brain metastases are similarly robust when derived from either 2-D spin echo or 3-D gradient echo MR images. However, features derived from 2-D spin echo images are generally less variable. Future work will compare predictive performance between these data sets.
Funding Support, Disclosures, and Conflict of Interest: Funding Support: MD Anderson - CCSG Radiation Oncology and Cancer Imaging Program Grant
Keywords
Not Applicable / None Entered.
Taxonomy
Contact Email
