A machine learning-based sonomics for prediction of thyroid nodule malignancies.

Area of Science:

• Radiology
• Medical Imaging
• Oncology

Background:

• Thyroid nodules are common, and distinguishing malignant from benign ones is crucial for patient management.
• Fine-needle aspiration (FNA) is the standard diagnostic method, but it can be invasive and sometimes inconclusive.
• There is a need for non-invasive biomarkers to improve diagnostic accuracy for thyroid nodules.

Purpose of the Study:

• To investigate the utility of ultrasound-derived radiomics features as non-invasive biomarkers for assessing thyroid nodule malignancy.
• To develop and evaluate machine learning models for classifying thyroid nodule malignancy based on these features.

Main Methods:

• Two hundred and ten patients underwent ultrasound-guided FNA for thyroid nodules.
• Radiomics features (intensity, shape, texture) were extracted from ultrasound images.
• Machine learning algorithms (LASSO, MRMR, Random Forests, XGBoost) were employed for feature selection and classification.

Main Results:

• Univariate analysis identified GLRLM-RLNU and GLZLM-GLNU as top predictors (AUC 0.67).
• Multivariate models achieved high performance, with AUCs up to 0.99 in the training set.
• The best performing model on the test set (XGBoost with MRMR and LASSO) showed an AUC of 0.95.

Conclusions:

• Ultrasound-derived radiomics features show promise as non-invasive biomarkers for thyroid nodule malignancy.
• Machine learning models utilizing these features can accurately predict malignancy, potentially improving diagnostic strategies.