Assessment of Clinical Radiosensitivity in Patients with Head-Neck Squamous Cell Carcinoma using Pre-treatment Quantitative Ultrasound Texture-Derivatives and Machine Learning Classifiers

Abstract

Quantititative ultrasound (QUS) provides information related to the tumor microstructure based on the tissue organization and elastic properties. To assess the efficacy of QUS in predicting recurrence risk and radiotherapy (RT) response at 3 months in patients with head-neck squamous cell carcinoma (HNSCC) treated with curative intent RT. Fifty-one patients with HNSCC with cervical lymph node (LN) were treated with 70 Gy/33 fractions (+/- chemotherapy) to the high-risk volumes according to standard institutional practice. The largest LN was scanned using a 6 MHz transducer before initiation of RT. QUS-parametric maps were generated from the radiofrequency data, from which texture features were extracted using grey-level co-occurrence matrices (GLCM). Further processing of texture images led to the generation of higher-order texture derivatives leading to a total of 105 parameters. After treatment completion, response assessment and follow up was done according to standard practice with clinical examination and imaging. Response at 3 months (primary and LN) and recurrence at any time period were considered as the end-points for this analysis. Machine learning algorithms based on linear discriminant (FLD), k-nearest-neighbors (KNN), support vector machine (SVM) were tested for developing classifier models with leave one out cross-validation. To avoid overfitting, we used a maximum of 3 features in model generation. The primary sites were oropharynx (34), hypopharynx (8), larynx (4), unknown primary (4), and oral cavity (1). Staining with p16 was positive in 35, negative in 2, and unknown in 13. Cisplatin was given concurrently in 37, Carboplatin in 5; and Cetuximab in 1. There were 21 complete responders (CR) and 30 partial responders (PR) at 3 months following RT completion. The best classifier performance was obtained using KNN (k = 2) with sensitivity, specificity, accuracy, and area under curve (AUC) values of 94%, 100%, 96% and 0.95 respectively. The inclusion of texture-derivatives led to significant improvement of the classifier accuracy compared to texture parameters alone (96% vs 86%). All the 3 selected features used in model construction were selected from higher-order texture-derivatives. With a median follow up of 30 months, recurrence was seen in 13 patients. The sensitivity, specificity, accuracy, and AUC were 82%, 85%, 83%, and 0.84 respectively, using SVM classifier to differentiate between the 2 groups (rec vs. no recurrence). This study shows QUS-texture derivatives coupled with machine learning can lead to the non-invasive determination of radiation response even before the initiation of RT with high accuracy. Similarly, patients with aggressive tumor biology with a higher risk of recurrent disease can be identified from pre-treatment QUS-radiomic signatures. Being an easily available, inexpensive, and portable imaging, QUS-radiomics can be potentially used to guide personalized RT in the future.