Ultrasound Image Classification of Thyroid Nodules Based on Deep Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodules are common, with a rising malignancy risk annually.
• Current diagnostic methods like fine needle aspiration (FNA) can be burdensome and invasive.
• Deep learning offers potential for non-invasive, accurate thyroid nodule diagnosis.

Purpose of the Study:

• To develop and validate a novel deep learning framework for precise prediction of benign and malignant thyroid nodules.
• To assess the performance of the proposed model using ultrasound image analysis.
• To explore the utility of Gradient-weighted Class Activation Mapping (Grad-CAM) in identifying diagnostic features.

Main Methods:

• A dataset of 508 thyroid ultrasound images was utilized for training and validation.
• A ResNet18 model, pre-trained on ImageNet, was employed for image recognition.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize and analyze image regions crucial for diagnosis.

Main Results:

• The deep learning model achieved high performance metrics, including an average Area Under Curve (AUC) of 0.997.
• Average accuracy reached 0.984, with an average recall of 0.978 and precision of 0.939.
• Grad-CAM analysis revealed distinct shape features in sensitive regions, aiding in differentiating benign from malignant nodules.

Conclusions:

• The proposed deep learning framework demonstrates high accuracy in distinguishing benign from malignant thyroid nodules using ultrasound images.
• The integration of Grad-CAM enhances diagnostic interpretability by highlighting key image features.
• This approach shows significant potential for improving thyroid nodule diagnosis, reducing the need for invasive procedures.