Prediction of Metastasis in Head and Neck Cancer from Computed Tomography Images

Abstract

The current medical method for determining whether the malignant tumor of the head and neck metastasizes to the lymph is to interpret the pathological section of the patient's lymph. This study proposes a support vector machine (SVM) based method Pred-Meta to predict metastasis of a malignant tumor from a patient's computed tomography (CT) image. Pred-Meta utilizes three feature types, including texture, morphology, and grayscale, and an optimal feature selection method cooperated with SVM. The data set consists of 75 samples from 70 patients in head and neck cancer provided by Taipei Veterans General Hospital of Taiwan with a record of lymphatic metastasis. Pred-Meta using leave-one-out cross-validation achieved 100% in predicting metastasis. The merit of the Pred-Meta method is its non-invasiveness and low cost. Auxiliary physicians screen out patients with high risk of diversion in the early stages to help plan treatment guidelines. The limitation of Pred-Meta suffers from the small number of training samples. It is expected that Pred-Meta would perform better in testing independent cohort when the number of training samples significantly increases.