Predicting Head and Neck Cancer Treatment Outcomes using Textural Feature Level Fusion of Quantitative Ultrasound Spectroscopic and Computed Tomography: A Machine Learning Approach

Predicting therapy response of Head & Neck (H&N) cancers prior to therapy initiation can be effective to increase the probability of pathologic complete response (pCR) and clinical response. Quantitative ultrasound spectroscopy (QUS) and treatment planning computed tomography (CT) are utilized to evaluate therapy response. Although analysis of cancerous cells on ultrasound (US) and CT images is not feasible due to their sub-resolution sizes, radiomics features can be extracted from these images to measure changes in biological conditions and cells’ micro-structures. Combination of CT and QUS data at feature level can generate features that are more informative. To this end, we proposed a technique to fuse radiomics feature of CT and QUS in order to generate more discriminative features. A fusion of radiomics features from CT and QUS was achieved using an autoencoder, followed by the application of an SVM classifier to distinguish between complete responders (CR) and partial responders (PR) among patients with H&N cancer. The proposed method could achieve to accuracy=71%, F1-score=69% and AUC=70% to classify H&N patient with responses CR and PR.