CQENet: A segmentation model for nasopharyngeal carcinoma based on confidence quantitative evaluation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of nasopharyngeal carcinoma (NPC) is crucial for effective radiotherapy.
• Current automatic segmentation methods struggle with tumor localization, boundary delineation, and confidence quantification.
• The 'black-box' nature of deep learning models limits clinical trust due to unquantified prediction confidence.

Purpose of the Study:

• To develop an automatic segmentation model for NPC that provides accurate tumor delineation and quantifies prediction confidence.
• To improve the clinical applicability of deep learning in NPC radiotherapy by addressing limitations in current methods.

Main Methods:

• Proposed a novel Confidence Quantitative Evaluation Network (CQENet) for NPC segmentation.
• Introduced a Confidence Assessment Module (CAM) to output segmentation results with associated confidence levels.
• Developed a Feature Adjustment Module (FAM) for precise tumor localization and extent determination.
• Incorporated a Variance Attention Mechanism (VAM) to enhance the delineation of blurred tumor boundaries.

Main Results:

• CQENet demonstrated superior performance compared to existing state-of-the-art models on a multicenter NPC dataset.
• The model successfully addressed challenges in tumor localization, boundary delineation, and confidence quantification.
• Experimental validation confirmed the effectiveness and clinical applicability of the proposed method.

Conclusions:

• CQENet offers a significant advancement in automatic NPC segmentation, providing both accurate results and confidence measures.
• The developed modules (CAM, FAM, VAM) effectively tackle key limitations of previous methods.
• The model shows considerable potential for enhancing NPC radiotherapy planning and clinical decision-making.