Whole-Body Radiomics For Prediction Of Treatment Failure In Cervical Cancer

To develop and validate a radiomics model for cervical cancer treatment outcomes that incorporates whole-body imaging biomarkers. We analyzed 96 cervical cancer patients treated definitively with chemoradiotherapy and intracavitary brachytherapy. All patients underwent pretreatment whole-body 18F-FDG PET/CT. A semi-automated approach combining seed-growing and manual contour review was employed to generate whole body muscle, bone, and fat segmentations on each PET/CT. Targets were deformably registered from each planning CT to the PET/CT. A total of 780 shape, intensity, and histogram radiomics features were extracted for targets, muscle, bone, and fat using open source software. Robustness of radiomics features was assessed via test/retest analysis using multiple same-day planning CT images, with significance indicated by p < 0.01 (F test). We trained and validated a Cox model of disease recurrence including both radiomic and clinical features (stage, tumor grade, histology, and baseline complete blood cell counts), using ridge regularization for feature selection and parameter fitting. Models were trained with leave-one-out cross validation, and evaluated for performance on both training and test sets using the C-index (C-index > 0.5 indicates nonzero predictive power). Thirty resamplings were performed with a 3:2 training:test set ratio. Bootstrap averages of the best C-indices for various models and training/test set combinations were calculated. A risk score for clinical use was created, based on the linear predictor from ridge regression, using normalized feature inputs. We found 459 significant features with 95% bootstrap confidence interval (CI) excluding 0, with 74 showing consistently all-positive or all-negative effects across all 30 bootstraps. Of the significant features, the most salient contributions were from CT-based texture and histogram features on the whole-body muscle and fat contours. Test/retest robustness analysis indicated that inter-patient feature variation was significantly greater than intra-patient feature variation across all feature types. The mean C-indices for the training set using only clinical features vs. clinical and radiomic features were 0.71 and 0.80, respectively. In the test set, the corresponding values were 0.48 and 0.55, indicating the inclusion of radiomic features added significant explanatory power to the model. Optimal stratification based on a threshold value of the risk score divided the data into high (N = 51) and low (N = 46) risk groups. The 2-year recurrence-free survival rate was 89.7% for the low-risk and 69.5% for the high-risk group, respectively (hazard ratio [95% CI]: 0.25 [0.10-0.62]; p = 0.006). Initial results indicate that incorporating whole-body radiomic features improves the performance of prognostic models as compared to using clinical features alone. Further work is needed to assess feature robustness and validate this radiomic modeling approach.