Thyroid nodules risk stratification through deep learning based on ultrasound images

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology Diagnostics

Background:

• Thyroid nodule risk stratification is crucial for clinical decision-making.
• The American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS) standardizes classification but is time-consuming and experience-dependent.
• Current deep learning (DL) models often lack clinical interpretability, hindering adoption.

Purpose of the Study:

• To develop an automated thyroid nodule risk stratification system using deep learning integrated with clinical experience.
• To create a DL model that is understandable and applicable in clinical practice.
• To enhance the accuracy and efficiency of thyroid nodule assessment.

Main Methods:

• Proposed a risk stratification network (RS-Net), an automatic hierarchical DL method.
• Incorporated ACR TI-RADS medical experience into the CNN-based classification system.
• Developed and evaluated the model using a dataset of 13,984 thyroid ultrasound images.

Main Results:

• Achieved 65% accuracy in risk stratification (TR1-TR5) with a mean absolute error (MAE) of 0.54.
• The MAE for point totals (0-13) was 1.67, with a Pearson's correlation of 0.84 compared to sonographers.
• Demonstrated high performance in benign vs. malignant classification (accuracy 88.0%, sensitivity 98.1%, specificity 79.1%).

Conclusions:

• The developed method automates thyroid nodule risk stratification, comparable to senior clinicians.
• The integration of clinical experience enhances DL model trust and clinical applicability.
• The system aids in improving diagnostic efficiency and reducing missed or misdiagnosed cases.