Predicting Recurrence for Patients with Head-Neck Squamous Cell Carcinoma Using Quantitative Ultrasound Based Radiomic Signatures Integrated with Machine Learning

Abstract

We have demonstrated the ability of quantitative ultrasound (QUS) in characterizing tumor biology in previous in vitro and in vivo studies. This aim of this project was to investigate the clinical utility of QUS as a biomarker for predicting recurrence in patients with node-positive head-neck squamous cell carcinoma (HNSCC) treated with radical radiotherapy (RT). Fifty patients with HNSCC were treated with RT (70 Gy/33 #) (+/- concomitant chemotherapy) were included in the current analysis. QUS data acquisition involved scanning nodes with a conventional-frequency device (7 MHz). Data were collected at the following time intervals relative to the start of treatment: week 0 (baseline), after week 1, and week 4. Acoustic backscatter features were quantified using spectral analysis of the radiofrequency data within a region of interest. Second order features included a set of 41 QUS radiomic features from a grey-level co-occurrence matrix. Imaging was acquired as part of the usual standard-of-care at 3 months after completing RT. Subsequently, patients were followed every 3 months for the initial two years and then every 6 months with clinical examination supplemented with imaging as indicated. Patients were categorized into two groups based on clinical outcomes (recurrence vs. no recurrence). Results were compared from 3 classification algorithms-Fisher linear discriminant, k-nearest-neighbors, and support vector machine (SVM). Three features used in each model were selected using a forward sequential selection method and validated using leave-one-out cross-validation. The median follow up for the entire group was 18 months. The most common primary site was oropharynx in 34 patients, followed by hypopharynx (7), larynx (4), unknown primary (4), and oral cavity (1). There were 16 complete responders (CR) and 34 partial responders (PR) at 3 months following completion of RT. Fourteen patients had recurrences: isolated regional (2), regional-distant (3), isolated distant (7), local-regional-distant (2), all belonging to the cohort of PR. The SVM classifier resulted in the best predictive performance (sensitivity, specificity, accuracy, and area under the curve values of 92%, 71%, 86%, and 71% respectively, for week 0). There was an improvement in classification accuracy at week 1 and week 4 using SVM classifier (Table 1). Most of the features selected for the recurrence classification were second-order QUS-texture features. Our preliminary results demonstrate encouraging results for QUS radiomic features as a potential biomarker for predicting patients at higher risk of disease recurrence. This is the first clinical report showing the ability of ultrasound to predict the outcomes even before the initiation of treatment in HNSCC.Abstract 2996; Table 1TimeModelSensitivitySpecificityAccuracyAUCWeek 0k-NN89%57%80%71%SVM92%71%86%71%Week 0 + ΔWeek 1k-NN89%57%80%71%SVM97%71%90%76%Week 0 + ΔWeek 4k-NN97%57%86%84%SVM100%71%92%87% Open table in a new tab