Symmetrical awareness network for cross-site ultrasound thyroid nodule segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning shows promise for ultrasound thyroid nodule segmentation.
• Challenges arise from limited annotations and domain shift in multi-site datasets.
• Poor generalization limits practical application of current deep learning methods.

Purpose of the Study:

• To develop an effective domain adaptation framework for improved ultrasound thyroid nodule segmentation.
• To enhance the generalization ability of deep neural networks across different medical imaging domains.
• To address the limitations of existing methods in handling multi-site, varied datasets.

Main Methods:

• A framework combining a bidirectional image translation module and two symmetrical segmentation modules.
• Mutual domain conversion using image translation to bridge source and target domains.
• Adversarial constraints and consistency loss for stable, efficient training and reduced domain gap.

Main Results:

• Achieved 96.22% for PA (Pixel Accuracy) and 87.06% for DSC (Dice Similarity Coefficient) on average.
• Demonstrated competitive cross-domain generalization ability compared to state-of-the-art methods.
• Validated performance on a multi-site ultrasound thyroid nodule dataset.

Conclusions:

• The proposed domain adaptation framework significantly improves ultrasound thyroid nodule segmentation.
• The method enhances deep learning model generalization, making it more robust to domain shifts.
• This work offers a promising solution for applying deep learning in real-world medical imaging scenarios.