• Joost J M van Griethuysen, LambregtsDoenja M JDMJhttp://orcid.org/0000-0003-2990-0099Department of Radiology, The Netherlands Cancer Institute, PO Box 90203, 1006 BE, Amsterdam, The Netherlands. d.lambregts@nki.nl., Stefano Trebeschi, Max J Lahaye, Frans C H Bakers, VliegenRoy F ARFADepartment of Radiology, Zuyderland Medical Center, Heerlen, The Netherlands., Geerard L Beets, Hugo J W L Aerts, and Beets-TanRegina G HRGHDepartment of Radiology, The Netherlands Cancer Institute, PO Box 90203, 1006 BE, Amsterdam, The Netherlands.GROW School for Oncology and Developmental Biology, Maastricht University, Maastricht, The Netherlands..
• Department of Radiology, The Netherlands Cancer Institute, PO Box 90203, 1006 BE, Amsterdam, The Netherlands.
• Abdom Radiol (NY). 2020 Mar 1; 45 (3): 632-643.

PurposeTo compare the performance of advanced radiomics analysis to morphological assessment by expert radiologists to predict a good or complete response to chemoradiotherapy in rectal cancer using baseline staging MRI.Materials And MethodsWe retrospectively assessed the primary staging MRIs [prior to chemoradiotherapy (CRT)] of 133 rectal cancer patients from 2 centers. First, two expert radiologists subjectively estimated the likelihood of achieving a "complete response" (ypT0) and "good response" (TRG 1-2), using a 5-point score (based on TN-stage, MRF/EMVI-status, size/signal/shape). Next, tumor volumes were segmented on high b value DWI (semi-automated, corrected by 2 non-expert and 2-expert readers, resulting in 5 segmentations), copied to the remaining sequences after which a total of 2505 radiomic features were extracted from T2W, low and high b value DWI and ADC. Stability of features for noise due to inter-reader and inter-scanner and protocol variations was assessed using intraclass correlation (ICC) and the Kruskal-Wallis test. Using data from center 1 (n = 86; training set), top 9 features were selected using minimum Redundancy Maximum Relevance and combined in a logistic regression model. Finally, diagnostic performance of the fitted models was assessed on data from center 2 (n = 47; validation set) and compared to the performance of the radiologists.ResultsThe Radiomic models resulted in AUCs of 0.69-0.79 (with similar results for the segmentations performed by expert/non-expert readers) to predict response, results similar to the morphologic prediction by the expert radiologists (AUC 0.67-0.83). Radiomics using semi-automatically generated segmentations (without manual input) did not result in significant predictive performance.ConclusionsRadiomics could predict response to therapy with comparable diagnostic performance as expert radiologists, regardless of whether image segmentation was performed by non-expert or expert readers, indicating that expert input is not required in order for the radiomics workflow to produce significant predictive performance.

25 keywords...

hide…