Improving Non-Invasive Prediction of Thyroid Nodule Malignancy: A Machine Learning-Based Clinical Approach.

Area of Science:

• Endocrinology
• Oncology
• Medical Imaging

Background:

• Thyroid cancer (TC) is a growing global health concern.
• Current diagnostic methods like ultrasound and fine-needle aspiration biopsy (FNAB) have limitations, often leading to unnecessary surgeries for benign nodules.
• Accurate preoperative risk stratification for thyroid nodules (TNs) remains a clinical challenge.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for classifying thyroid nodules as malignant or benign.
• To utilize a select set of ultrasonographic features for non-invasive risk stratification.
• To assess the potential of ML in reducing unnecessary thyroidectomies.

Main Methods:

• Retrospective analysis of data from 5928 patients who underwent thyroidectomy.
• Inclusion of five key ultrasonographic features: hypoechogenicity, microcalcifications, shape, irregular margins, and vascularity.
• Training and evaluation of five ML models: Random Forest, Logistic Regression, Multilayer Perceptron (MLP), Gradient Boosting Machines, and Decision Tree.

Main Results:

• The Random Forest model demonstrated the highest performance with an accuracy of 0.905 and a specificity of 0.939.
• Nodule vascularity was the most significant predictor, followed by microcalcifications, irregular margins, and hypoechogenicity.
• The models showed potential in identifying benign nodules but had a modest recall (0.616) for malignancy detection.

Conclusions:

• Machine learning models utilizing specific ultrasonographic features can aid in the non-invasive identification of benign thyroid nodules.
• These ML models may help reduce the number of unnecessary surgical procedures.
• Further improvements in sensitivity require integrating additional imaging parameters and cytological data for reliable standalone cancer detection.