Is the Promise of Diffusion-weighed (DWI) MRI based Therapy Response Prediction in Advanced Cervical Cancer Ready for Prime Time in Standard Community Settings? The Issue of Standardized Postprocessing

Abstract

Functional imaging (DWI) has shown promise in predicting response and tumor control in cervical cancer treated with radiation/chemotherapy, but clinical translation has been limited by complexities in image registration, tumor delineation on apparent diffusion coefficient (ADC) maps, and artifacts. Two methods, scanner-generated auto-coregistration (Coaut) and more resource-intense individualized coregistration (Coind) may influence DWI results. Our purpose was to compare both methods' consistency and accuracy for tumor response prediction and assess potential pitfalls. We analyzed 21 FIGO stage IB2-IVA cervical cancer patients enrolled in a prospective imaging trial who were treated with concurrent chemoradiation. Coaut vs. Coind were studied in each with 3 serial MRIs, pre-therapy (pre), early-therapy (2-wk) and mid-therapy (5-wk) obtained. ADC data sets contained T2- weighted (T2w)-based whole-tumor contouring and were linked to 1-month post-therapy tumor response. For Coind, the scanner-DICOM-generated Coaut were adjusted individually to account for distortion and shifts of the tumor region between T2w and DWI. Intensity histogram tumor features, mean, coefficient of variation, skewness, and kurtosis were extracted from the Coaut and Coind MRI ADC images at each time point and correlated with treatment response. Pairwise consistency of features was evaluated with signed rank testing. Predicted response classification was compared by ROC area-under-the-curve (AUC) and rank sum testing. Pre and 2-wk mean ADC were consistent (<20% variability, p>0.34) between Coaut and Coind in most patients (20/21 each); 5-wk mean ADC was less consistent (18/21). Inconsistencies were due to tumor distortion/shifts (4 patients) and artifact (1). Coind-based pre and 5-wk mean ADC predicted response (AUC 0.81 and 0.82, p = 0.021 and 0.017, respectively); Coaut-based pre mean ADC showed similar performance (AUC = 0.81, p = 0.021) but not 5-wk mean ADC (AUC 0.65, p = NS). Coind-based decrease in 2-wk ADC skewness best predicted response (AUC = 0.84, p = 0.013), outperforming Coaut (AUC 0.71, p = NS). Inconsistencies between scanner automated and individualized co-registration were small in the vast majority of cases for whole-tumor based pre-therapy mean ADC, which predicted response with both methods. Major ADC distortion and ADC- T2w misalignments that were easily identifiable benefited from additional individualized co-registration adjustment. Pre-therapy DWI MRI for response prediction in cervical cancer is readily exportable to community settings using scanner based post-processing. This capability may broaden the application of DWI for treatment of cervical cancer in the wider community. Supported by R01CA155454.