Development of a Radiomic Signature Derived from Apparent Diffusion Coefficient Maps to Predict Outcomes in Cervical Cancer

Abstract

<h3>Purpose/Objective(s)</h3> Previous studies have suggested that apparent diffusion coefficient (ADC) values from pretreatment MR are associated with outcomes after definitive chemoradiation in patients with cervical cancer. The purpose of this study was to determine whether radiomic features derived from ADC maps on pretreatment diffusion-weighted MRI are predictive of patient outcomes, and prioritize those individual features into a predictive signature. <h3>Materials/Methods</h3> ADC maps were generated from pretreatment pelvic diffusion-weighted MRIs acquired on a 1.5T MRI simulator for twenty-six patients prior to definitive radiotherapy (RT). All patients were treated with curative intent chemoradiation, including external beam RT and image guided brachytherapy as per the standard of care. Median post therapy clinical follow up time was 67 months, and treatment outcomes were approximated by the results of 3-month post therapy FDG-PET. The clinical endpoint was recurrence-free survival (RFS), defined as death or recurrence of cancer. High risk regions, visually identified as areas of ADC hypointensity, were manually contoured. Quantitative radiomic features characterizing intensity, shape, texture and multiscale wavelet transformations of contoured regions of interest were identified using an unsupervised learning approach. A radiomic signature for RFS was generated using a multivariate Cox model that included the best performing feature from each of the four categories. <h3>Results</h3> Our model identified high and low risk groups based on a radiomic signature that stratified patients by RFS (hazard ratio 4.5, p=0.0016). The features chosen for the signature included the median of voxel intensity, joint entropy of the gray level co-occurrence matrix (GLCM), elongation of the region of interest, and LHH wavelet transform of intensity skewness. The mean receiver-operator curve on 5-fold cross-validation had an area under the curve of 0.83 +/- 0.16. <h3>Conclusion</h3> In this pilot study, we have developed a radiomic signature derived from pretreatment ADC maps of cervix tumors that successfully stratifies this group of uniformly treated patients by RFS. Future work will seek to validate our model on a larger independent data set to determine its predictive value as well as correlate these data with radiomic features from FDG-PET/CT images and RNA-seq data from patient tumor samples.