Machine Learning Assisted Doppler Features for Enhancing Thyroid Cancer Diagnosis: A Multi-Cohort Study

Area of Science:

• Radiology and Artificial Intelligence
• Medical Imaging Analysis
• Oncology Diagnostics

Background:

• Thyroid cancer classification relies on ultrasound (US) imaging.
• Improving diagnostic accuracy for thyroid nodules (TNs) is crucial.
• Machine learning (ML) offers potential for enhanced image analysis.

Purpose of the Study:

• To develop and evaluate an artificial neural network (ANN) model for thyroid cancer classification.
• To integrate color Doppler ultrasound (CDUS) features with gray-scale US features for improved diagnostic performance.
• To compare the diagnostic performance of the ML model against radiologists.

Main Methods:

• A retrospective study included 712 thyroid nodules (TNs) from internal and external datasets.
• An ANN model (TDUS-Net) was developed using quantitative CDUS features and American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS) gray-scale US features.
• The diagnostic performance of TDUS-Net was compared with another ANN model using only gray-scale features (TUS-Net) and with human radiologists.

Main Results:

• TDUS-Net demonstrated a higher area under the curve (AUC) of 0.898 compared to TUS-Net's AUC of 0.881 in internal tests.
• In external tests, TDUS-Net achieved a superior AUC of 0.925, significantly outperforming radiologists with an AUC of 0.810.
• The study highlights the effectiveness of combining different ultrasound feature types for classification.

Conclusions:

• Integrating gray-scale US and CDUS features via ML enhances thyroid nodule classification accuracy.
• The developed TDUS-Net model shows potential to match or exceed the diagnostic capabilities of experienced radiologists.
• This approach offers a promising tool for improving the early and accurate diagnosis of thyroid cancer.