Speaker: Tahir I. Yusafaly

Department of Radiation Medicine and Applied Sciences
University of California, San Diego
La Jolla, California


Title: Whole-Body PET/CT Radiomics to Predict Treatment Failure in Cervical Cancer Patients


Abstract: Radiomics is the application of machine learning methods to extract clinically useful information from medical images. Radiomic models have been shown to improve the ability to predict clinical outcomes, whether toxicity or tumor control, in cancer patients undergoing standard-of-care treatment regimens. The most well-developed applications of radiomics focus on features derived from anatomy-based imaging localized in the immediate vicinity of the tumor target. In contrast, relatively little analysis has been done using features derived from functional imaging techniques, and even less has been done interrogating the properties of ‘off-target’ whole-body imaging features. 

The purpose of this work, therefore, was to develop and validate a model to predict treatment failure using radiomic features derived from pre-treatment ‘whole-body’ positron emission tomography/computed tomography (PET/CT) images. We analyzed 127 cervical cancer patients treated definitively with chemoradiotherapy and intracavitary brachytherapy. A semi-automated approach combining seed-growing and manual contouring generated three whole-body (muscle, bone, and fat) and two target (gross target volume (GTV) and planning target volume (PTV)) segmentations on each PET/CT. Ninety-five patients from the cohort were used to train a Cox model of disease recurrence including both radiomic and clinical features (age, stage, tumor grade, histology, and baseline complete blood cell counts), using bagging and split-sample-validation for feature selection. The resulting model was tested on the remaining 32 patients, by calculating a risk score based on Cox regression and evaluating the C-index (C-index > 0.5 indicates predictive power).

Optimal performance was seen in a Cox model including one clinical feature (stage), two PET-based features of the GTV, and one CT-based feature of the whole-body bone segmentation. A tentative interpretation of these findings suggests that the PET-based GTV metrics are proxies for metabolic heterogeneity in the tumor microenvironment, while the CT-based bone metric is a proxy for tumor-associated cachexia. These results suggest that incorporating whole-body and functional imaging radiomics improves prognostic performance compared to using clinical features alone. However, further work is necessary to prospectively test these models in larger, multi-institutional cohorts.