MULTIMODAL MACHINE LEARNING FOR DIFFERENTIAL DIAGNOSIS OF CHRONIC OTALGIA

Abstract

Lymph node metastasis is a key prognostic factor and is also a significant factor in the treatment planning, staging, risk of recurrence and prognosis of head and neck cancers. The common radiological assessment relies mostly on the morphology and size of the nodes, necrosis and enhancement patterns, but these are not always sufficient to represent the underlying heterogeneity of primary tumors and metastatic lymph nodes. In this paper, we introduce a machine learning-based framework for lymph node metastasis prediction for head and neck cancer patients by leveraging quantitative imaging-based features from medical images. Radiomic parameters, such as shape, intensity, texture, and other higher order imaging parameters, are extracted from the tumor and nodal region to describe the heterogeneity of the disease beyond visual inspection. Once the data has been preprocessed, feature selection will be carried out to remove redundant features and to discover the most discriminative features. Then, multiple machine learning classifiers are trained and evaluated for the prediction of metastatic lymph nodes. The accuracy, sensitivity, specificity, precision, F1-score and area under the receiver operating characteristic curve are used to assess model performance. The proposed approach will seek to aid the clinician with a non-invasive, reproducible and data-driven algorithm to stratify patients for higher risk of nodal involvement. The results can help to enhance pre-treatment risk stratification, and potentially aid in individualized therapeutic choice for patients with head and neck cancers in the future.