CT imaging during treatment improves radiomic predictions for patients with locally advanced head and neck cancer

Introduction: Radiomics applies machine learning algorithms to characterize the tumor phenotype and to predict clinical outcome based on quantitative imaging data. It has been applied using pre-treatment computed tomography (CT) scans but only few studies have assessed radiomics on imaging acquired during radiotherapy. Therefore, we compared the performance of radiomic models based on the pre-treatment CT with that based on CT imaging during treatment.
Material/Methods: Two datasets of patients with advanced stage head and neck squamous cell carcinoma (HNSCC) were used as an exploratory and a validation cohort (47 and 30 patients, respectively). All patients received primary radio-chemotherapy (RCT) and underwent a non-contrast-enhanced CT scan pre-treatment and in week 2 (W2) of treatment. 1610 image features were extracted from the gross tumour volume, delineated on the baseline CT and the W2 CT. Radiomic models were built to predict loco-regional tumour control (LRC). Different feature selection methods (mutual information maximization (MIM), random forest variable importance (RFVI)) and learning algorithms (Cox regression (COX), random forest (RF)) were evaluated using the concordance index (CI) as performance measure.
Results: On the W2 CT both FS methods combined with the RF algorithm achieved a higher performance (CI=0.71) than on the baseline CT (CI<0.65), which was also observed using the Cox regression model (W2 CT: CI=0.66, baseline CT: CI=0.51).
Conclusions: CT scans from the second week of RCT for patients with locally advanced HNSCC improved the performance of radiomic prediction models compared to baseline CT scans. The incorporation of during-treatment imaging is a promising way to improve radiomic models for clinical treatment adaption.