Convolutional Neural Network for Predicting Thyroid Cancer Based on Ultrasound Elastography Image of Perinodular...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Thyroid nodules (TNs) are common, and accurate cancer prediction is crucial.
• Current diagnostic methods can be limited, necessitating improved imaging analysis.
• Deep learning offers potential for enhanced diagnostic capabilities in radiology.

Purpose of the Study:

• To develop and evaluate deep learning models for thyroid cancer prediction using ultrasound (US) and shear-wave elastography (SWE) images.
• To compare the performance of single-image and fused-image convolutional neural network (CNN) models.
• To assess the impact of perinodular SWE image characteristics on diagnostic accuracy.

Main Methods:

• Retrospective study of 1747 thyroid nodules (TR4) from 1582 patients.
• Development of 7 single-image and 6 fused-image CNN models (RestNet18) using US and various SWE image parameters (perinodular regions, ROI).
• Evaluation of models on a training (1247 TNs) and validation (500 TNs) cohort for thyroid cancer prediction.

Main Results:

• The US + 2.0 mm SWE image CNN model achieved the highest area under the curve (AUC) for nodules >10 mm (0.95 training, 0.92 validation).
• The US + 1.0 mm SWE image CNN model demonstrated the highest AUC for nodules ≤10 mm (0.95 training, 0.92 validation).
• Fused CNN models integrating SWE segmentation images and US images significantly improved the radiological diagnostic accuracy for thyroid cancer.

Conclusions:

• Deep learning models combining US and SWE imaging data enhance the prediction of thyroid cancer.
• The fusion of perinodular SWE image data with US images offers superior diagnostic performance compared to US alone.
• These findings suggest a valuable role for AI-driven image analysis in improving thyroid nodule risk stratification.