Interpretable Machine Learning Model Using Digitized US Features for Classifying Complex Thyroid Nodules.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Thyroid nodules are common, requiring accurate classification for appropriate management.
• Distinguishing between benign and malignant thyroid nodules, especially complex cases, remains a clinical challenge.
• Current diagnostic methods may benefit from advanced computational approaches for improved accuracy.

Purpose of the Study:

• To develop a digitized, interpretable machine learning classification model for thyroid nodules.
• To accurately recognize complex thyroid nodules and efficiently diagnose conventional ones.
• To integrate digitized ultrasound features into a white-box framework for enhanced classification.

Main Methods:

• Retrospective collection of thyroid ultrasound images from seven Chinese medical centers (2011-2021).
• Development of UltraMC, a two-layer interpretable classification model with front-end and back-end networks.
• Evaluation of UltraMC using accuracy, sensitivity, specificity, and ROC curves.

Main Results:

• The dataset comprised 73,826 patients; the front-end network achieved 92.9% accuracy for conventional nodules.
• The back-end network achieved 88.5% accuracy for mummified thyroid nodules (MTNs).
• Overall diagnostic accuracy of UltraMC for MTN classification was 91.8%, with high AUC values.

Conclusions:

• The two-layer interpretable classification model (UltraMC) demonstrates high diagnostic accuracy for both conventional and mummified thyroid nodules.
• Digitized ultrasound features within a white-box framework effectively support the classification of complex thyroid nodules.
• This approach offers a promising tool for improving the diagnostic capabilities in thyroid nodule assessment.