Using Radiomics to Improve the Diagnostic Accuracy of Indeterminate Residual Primary Disease on Restaging PET/CT Imaging Following Radiation Therapy for Head and Neck Cancers

Abstract

<h3>Purpose/Objective(s)</h3> Identifying residual disease (RD) in the primary site of head and neck cancer (HNC) patients treated with radiotherapy (RT) on the 3-month post RT PET/CT scans can be challenging. Radiologists have to account for RT-related inflammation versus RD. Accurate identification is key to determine the need for morbid salvage versus close observation. In this study, we hypothesized a clinical/radiomics model could complement a radiologist's interpretation in these uncertain cases. <h3>Materials/Methods</h3> We identified 182 patients with HNC treated at our institution from 2016-2020. We reviewed the 3-month post-RT PET/CT scan interpretations to obtain the radiologists' RD assessment. We classified assessments as "certain" or "uncertain." "Certain" assessments included words like "complete metabolic response," "no RD," "RD present" while "uncertain" assessments provided a differential. For uncertain cases, predictions were inferred as the most likely diagnosis ("likely radiation changes but can't rule out RD" was no RD present). If no clarification was provided, then the prediction was marked as RD present. We extracted radiomics features from the CT and PET modalities using open-source software. We gathered clinical data such as HPV status, T stage, etc. Feature reduction was performed by assessing single feature performance on a logistic regression model. Features with high correlation (>0.9) measured using Pearson coefficient were discarded. Recursive feature elimination was performed on the remaining features (only 5-10 features were allowed in the final model). We trained an explainable boosting machine model using the "certain" data and tested with "uncertain" data. Ground truth was obtained from follow-up imaging or pathology. <h3>Results</h3> Ninety-six assessments were "certain" and 86 were "uncertain." RD was present in 10 and 9 cases, respectively. The performances for the radiologists and our models are summarized in Table 1. The Clinical/CT radiomics model reduced the number of false positives from 22 to 8 at the cost of only one additional false negative. The ensemble of our CT and PET models did not improve performance. Table 1 <h3>Conclusion</h3> Even expert radiologists may struggle with identifying RD given post-RT inflammatory changes. Our Clinical/CT radiomics model greatly improved specificity with only a minor drop in sensitivity in this uncertain cohort. Machine learning models may have utility in complementing clinical experience to improve diagnostic accuracy for this challenging scenario.