TEM-Net: A Tri-Channel Edge-Aware Multi-Scale Network for Thyroid Nodule Segmentation in Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Thyroid nodule segmentation in ultrasound is vital for assessing features like contour and margins.
• Challenges include low contrast, blurred boundaries, and variations in nodule size and shape.
• Accurate segmentation is essential for computer-aided diagnosis systems.

Purpose of the Study:

• To develop an advanced deep learning model for accurate thyroid nodule segmentation in ultrasound images.
• To address the limitations of existing methods in handling low contrast and boundary ambiguity.
• To improve the reliability of computer-aided assessment of thyroid nodules.

Main Methods:

• Proposed TEM-Net, a Tri-Channel Edge-Aware Multi-Scale Network.
• Utilized a tri-channel input: grayscale, contrast-enhanced, and gradient-magnitude images.
• Incorporated an Edge-Guided Feature Amplification (EGFA) module and a Multi-Focus Cross-Scale Attention Refinement (MF-CAR) module.

Main Results:

• TEM-Net achieved competitive performance on TN3K and DDTI datasets.
• Mean Dice scores of 0.8822 (TN3K) and 0.9066 (DDTI).
• Mean IoU scores of 0.7893 (TN3K) and 0.8291 (DDTI).

Conclusions:

• TEM-Net effectively segments thyroid nodules in challenging ultrasound images.
• The proposed network architecture enhances boundary detection and feature fusion.
• Demonstrated superior performance compared to existing segmentation methods.