Spatial Prior-Guided Dual-Path Network for Thyroid Nodule Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate thyroid nodule segmentation in ultrasound is crucial for diagnosis but challenging due to image quality and anatomical complexity.
• Current deep learning methods often neglect thyroid anatomical priors, leading to misclassification of non-thyroid tissues.

Purpose of the Study:

• To develop a deep learning network that integrates anatomical prior knowledge for improved thyroid nodule segmentation.
• To enhance segmentation accuracy and boundary delineation by leveraging both global context and local features.

Main Methods:

• Proposed a Spatial Prior-Guided Dual-Path Network with a prior-aware encoder for anatomical structures and a heterogeneous encoder for multi-scale features.
• Introduced a CrossBlock module combining cross-attention and mixed-scale convolutions for global and local feature extraction.
• Utilized a dual-decoder architecture for thyroid region prior learning and nodule segmentation, with hierarchical feature injection for anatomical guidance.

Main Results:

• The proposed network demonstrated superior performance over state-of-the-art methods on TN3K and MTNS datasets.
• Achieved significant improvements in boundary precision and localization accuracy for thyroid nodule segmentation.
• The method showed practical value for preoperative planning and clinical decision-making.

Conclusions:

• The Spatial Prior-Guided Dual-Path Network effectively addresses limitations of existing methods by integrating anatomical priors.
• The network enhances segmentation accuracy, particularly in boundary delineation and localization.
• This approach offers a valuable tool for improving diagnostic accuracy in thyroid ultrasound imaging.