Differentiating malignant thyroid nodule with statistical classifiers based on demographic and ultrasound features

The present work aims to assess the role of statistical classifiers in increasing the diagnostic accuracy when differentiating between benign and malignant thyroid nodules. The classifiers considered were based on combinations of both demographic and ultrasound data. Two feature selection procedures were considered, along with three different classification methods: Random Forest, Support Vector Machine and Logistic Regression. The best results were obtained with the latter, for which the 95% confidence intervals for the area under the ROC curve, the sensitivity and the specificity were [0.75, 0.80], [76%, 83%] and [71%, 77%], respectively. The most relevant features were a description of the contour of the nodule, presence of halo, internal composition, sphericity and echogenicity, as well as the age of the participants.