Ultrasonic thyroid nodule detection method based on U-Net network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Current thyroid nodule detection methods are time-consuming.
• Feature extraction in existing methods presents challenges.
• Accurate segmentation is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To develop an efficient and accurate thyroid nodule detection system.
• To improve upon existing computed-aided diagnosis (CAD) methods.
• To leverage deep learning for enhanced thyroid nodule segmentation.

Main Methods:

• A mark-guided ultrasound deep network segmentation model based on U-Net was proposed.
• The U-Net model's performance was compared against VGG19, Inception V3, and DenseNet 161.
• Evaluation metrics included segmentation accuracy, edge precision, and network operation time.

Main Results:

• The U-Net model achieved near 100% overlap with manually segmented nodule areas.
• Segmentation accuracy reached 0.9785, outperforming other networks by approximately 3%.
• U-Net segmentation results closely matched manual delineations.

Conclusions:

• The proposed U-Net based segmentation significantly enhances thyroid nodule detection accuracy.
• The model performs well even with small training datasets.
• This approach offers a valuable reference for clinical diagnosis and treatment of thyroid nodules.