SECP-Net: SE-Connection Pyramid Network for Segmentation of Organs at Risk with Nasopharyngeal Carcinoma

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of organs at risk (OAR) in computed tomography (CT) images is crucial for nasopharyngeal carcinoma (NPC) treatment planning.
• Deep learning models like U-Net are widely used but struggle with variable OAR sizes and small volumes in NPC segmentation.
• Traditional methods often fail due to insufficient global and multi-size information utilization.

Purpose of the Study:

• To propose a novel deep learning model, the SE-Connection Pyramid Network (SECP-Net), for enhanced OAR segmentation in NPC.
• To address the limitations of existing models in handling variable and small OAR sizes.
• To improve the accuracy and efficiency of automatic medical image segmentation for NPC patients.

Main Methods:

• Developed SECP-Net incorporating an SE-connection module and a pyramid structure to capture global and multi-size features.
• Implemented an auto-context cascaded structure for refining segmentation outcomes.
• Conducted comparative experiments against recent methods using private head and neck CT and public liver datasets.

Main Results:

• SECP-Net demonstrated superior performance in OAR segmentation compared to existing methods.
• The model achieved state-of-the-art (SOTA) results on both the private and public datasets.
• Performance was evaluated using Dice and Jaccard similarity metrics via five-fold cross-validation.

Conclusions:

• SECP-Net effectively addresses the challenges of segmenting variable and small OARs in NPC CT images.
• The proposed architecture significantly improves segmentation accuracy, offering a valuable tool for clinical applications.
• SECP-Net represents a significant advancement in deep learning-based medical image segmentation for cancer treatment.