A machine-learning algorithm for distinguishing malignant from benign indeterminate thyroid nodules using ultrasound...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology Diagnostics

Background:

• Ultrasound-guided fine needle aspiration (FNA) is standard for thyroid nodule evaluation.
• Up to 30% of FNA results are indeterminate, necessitating further diagnostic procedures.
• Improved methods are needed for accurate classification of indeterminate thyroid nodules.

Purpose of the Study:

• To develop and evaluate a machine learning model for classifying indeterminate thyroid nodules.
• To enhance the accuracy of thyroid cancer diagnosis using radiomic features from ultrasound images.
• To compare the performance of the machine learning model against a commercial molecular testing platform.

Main Methods:

• Collected and analyzed 1052 ultrasound images from 302 thyroid nodules across two institutions.
• Manually annotated nodule margins and performed computerized radiomic texture analysis.
• Utilized a Bayesian artificial neural network classifier with stepwise feature selection and cross-validation; tested on an independent dataset.

Main Results:

• The machine learning model achieved an area under the curve (AUC) of 0.75 for malignant versus benign nodules and 0.67 for indeterminate malignant versus indeterminate benign nodules in training/validation.
• On an independent test set, the algorithm distinguished indeterminate nodules with an AUC of 0.88.
• This performance was superior to a commercial molecular testing platform, which yielded an AUC of 0.81.

Conclusions:

• Machine learning analysis of computer-extracted texture features from grayscale ultrasound images shows significant potential.
• The developed model effectively classifies indeterminate thyroid nodules based on surgical pathology.
• This AI-driven approach offers a promising, non-invasive tool for improving thyroid cancer diagnosis.