Thyroid Nodule Malignancy Risk Stratification Using a Convolutional Neural Network.

Area of Science:

• Artificial Intelligence in Radiology
• Medical Imaging Analysis
• Oncology

Background:

• Thyroid nodules are common, and accurate risk stratification is crucial for patient management.
• Traditional methods like ACR TIRADS assist in classifying nodules but can have limitations.
• Developing advanced tools to improve diagnostic accuracy for thyroid nodules is essential.

Purpose of the Study:

• To evaluate the performance of convolutional neural networks (CNNs) in stratifying the malignant potential of thyroid nodules.
• To compare CNN-based risk scores with the established American College of Radiology Thyroid Imaging Reporting and Data System (ACR TIRADS) classification.
• To assess the utility of CNNs as an adjunct tool for radiologists.

Main Methods:

• A dataset of 651 pathology-proven thyroid nodules was analyzed using grayscale ultrasound images.
• A CNN classifier, specifically a MobileNet ensemble, was developed and validated using a nested double cross-validation scheme.
• CNN-generated malignancy scores were compared with ACR TIRADS classifications for all nodules.

Main Results:

• The best-performing MobileNet CNN ensemble achieved an area under the curve of 0.86 (95% CI, 0.83-0.90).
• CNN risk strata showed distinct malignancy rates (81.4% for highest, 5.9% for lowest).
• CNN scores correlated well with ACR TIRADS levels, showing a gradient of malignancy risk.

Conclusions:

• CNNs can be trained to generate accurate malignancy risk scores for thyroid nodules.
• These AI models can serve as an adjunct tool to traditional guidelines like ACR TIRADS.
• CNNs can assist radiologists in identifying patients who may require further histopathologic workup.