Video-based AI module with raw-scale and ROI-scale information for thyroid nodule diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Thyroid Nodule Diagnosis

Background:

• Ultrasound is a key method for thyroid lesion detection.
• Inaccurate diagnoses can lead to delayed treatment or unnecessary biopsies.
• There is a need for improved precision in thyroid ultrasound diagnosis.

Purpose of the Study:

• To develop an artificial intelligence (AI) model for enhanced thyroid ultrasound diagnosis.
• To increase diagnostic precision and reduce the rate of unnecessary biopsy procedures.

Main Methods:

• Collected ultrasound recordings from 672 patients (845 nodules) across two hospitals.
• Developed and tested six AI model variants using different video feature strategies and Region of Interest (ROI) scale information.
• Evaluated model performance on internal and external test sets, comparing against experienced radiologists.

Main Results:

• The dual-stream AI model, incorporating raw-scale, ROI-scale streams, and time-dimensional convolution, achieved the highest performance.
• Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.969 and 92.6% accuracy on the internal test set.
• Demonstrated superior performance compared to other AI variants and experienced radiologists, with external test AUROC of 0.931.

Conclusions:

• The proposed dual-stream AI model significantly enhances thyroid ultrasound diagnostic accuracy.
• Integration of ROI scale information and time-dimensional convolution improves diagnostic performance.
• This AI approach shows potential to reduce unnecessary thyroid nodule biopsies.