Thyroid Nodule Region Estimation using Auto-Regressive Modelling and Machine Learning

Abstract

Ultrasound (US) imaging is used for the diagnosis and also evaluation of thyroid nodules. A Thyroid Imaging Reporting and Data System (TIRADS) is used for the risk stratification of thyroid nodules through US images. The composition of thyroid nodules plays an important role in the risk-stratification process. The percentages of cystic and solid components in a thyroid nodule are one of the features that are can be indicative of the risk of malignancy. In this work, we attempt to classify and estimate solid and cystic regions within nodules. 20x20 texture patches were extracted from solid and cystic regions and converted into signals. These signals are decomposed into low, mid, and high-frequency bands using Continuous Wavelet Transform (CWT). A total of 36 features were extracted from the decomposed signals using Auto- Regressive Modeling. The features were fed into three different Machine Learning (ML) algorithms (Artificial Neural Networks, K-Nearest Neighbors, and Random Forest Classifier) to provide us with a classification of solid versus cystic regions in thyroid nodule US images. The Random Forest Classifier obtained an Accuracy, Sensitivity, and Specificity of 90.41%, 99% and 91% respectively which was the highest among the three chosen ML algorithms. Additionally, the output from the classification phase was also be used to determine the percentage of cystic and solid regions with a given thyroid nodule US image.

Keywords

• ultrasound imaging
• thyroid nodules
• feature extraction
• region estimation
• machine learning
• classification