PCG-net: feature adaptive deep learning for automated head and neck organs-at-risk segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of Head and Neck (HN) Organs-At-Risks (OARs) is crucial for effective Head and Neck Cancer (HNC) radiation therapy planning.
• Manual OAR segmentation is time-consuming and subjective, driving the need for automated deep learning solutions.
• Segmenting small HN OARs like the optic chiasm and optic nerve remains a challenge for current deep learning models.

Purpose of the Study:

• To introduce PCG-Net, a novel parallel network architecture for improved HN OARs segmentation.
• To enhance feature fusion and capture both local and global contextual information for more accurate segmentation.
• To evaluate the effectiveness and robustness of PCG-Net in downstream segmentation tasks.

Main Methods:

• Developed PCG-Net, a parallel network integrating Convolutional Neural Networks (CNNs) and a Gate-Axial-Transformer (GAT).
• Incorporated a Cascade Graph Module (CGM) to improve feature fusion via message-passing and aggregation.
• Conducted extensive experiments to validate PCG-Net's performance across three different tasks.

Main Results:

• PCG-Net demonstrated superior performance compared to existing methods in HN OARs segmentation.
• The proposed model significantly improved segmentation accuracy, particularly for small-sized OARs.
• Experimental results confirmed the robustness and effectiveness of PCG-Net.

Conclusions:

• PCG-Net successfully integrates local and global information processing for accurate HN OAR segmentation.
• The CGM module effectively enhances feature fusion, leading to superior segmentation outcomes.
• PCG-Net shows significant promise for advancing HNC treatment planning through improved OAR segmentation.