Head-and-Neck Organs Segmentation in CT Based on Spatial Prior and Shape Description.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate delineation of organs at risk (OARs) is crucial for effective head and neck cancer radiotherapy.
• Current deep learning methods struggle with segmenting small OARs due to their quantity, distribution, and shape complexity.

Purpose of the Study:

• To develop a novel deep learning framework for enhanced segmentation of small OARs in head and neck cancer.
• To improve the accuracy of radiotherapy planning by addressing limitations in current segmentation techniques.

Main Methods:

• Proposed a spatial guidance network (SG-Net) to generate spatial guidance maps (SGMs), emphasizing organ boundaries and relationships.
• Introduced a deep shape description module (DSDM) to extract and integrate organ-specific shape features for small organs.
• Implemented a regularization term to preserve shape details and reduce over-smoothing in segmentation probability maps.

Main Results:

• The novel framework significantly improved the segmentation accuracy of small organs at risk.
• Segmentation accuracy for large organs was maintained.
• The proposed method demonstrated superior performance compared to state-of-the-art techniques for small organ segmentation.

Conclusions:

• The developed deep learning framework effectively addresses the challenge of segmenting small OARs in head and neck cancer.
• This approach enhances radiotherapy precision by improving the accuracy of critical structure delineation.
• The method offers a significant advancement over existing deep learning segmentation techniques for this clinical application.