Multi-view attribute learning and context relationship encoding enhanced segmentation of lung tumors from CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Graph convolutional neural networks (GCNs) offer flexibility in medical image segmentation by representing image regions as nodes and knowledge propagation via edges.
• Existing GCN methods often fail to fully leverage diverse image node attributes and their inter-attribute context relationships.

Purpose of the Study:

• To propose a novel medical image segmentation method, MNSeg, that enhances attribute learning and context relationship encoding.
• To improve segmentation performance by effectively integrating multi-view node attributes and contextual information.

Main Methods:

• MNSeg employs a GCN-based multi-view image node attribute learning (MAL) module to integrate attributes from multiple similarity views.
• A transformer-based context relationship encoding (CRE) strategy is utilized to capture and propagate attribute relationships across image nodes.
• An attention at attribute category level (ACA) module adaptively learns the importance of attribute categories for discrimination and fusion.

Main Results:

• MNSeg demonstrated superior performance over existing methods on public lung tumor CT and in-house NSCLC datasets, showing improved spatial overlap and shape similarity.
• Ablation studies confirmed the effectiveness of the MAL, CRE, and ACA modules.
• Consistent performance improvements across different 3D segmentation backbones validated MNSeg's generalization ability.

Conclusions:

• MNSeg offers a significant advancement in medical image segmentation by effectively utilizing multi-view attributes and context relationships.
• The proposed method shows strong potential for accurate and robust segmentation of challenging medical images, including lung tumors and NSCLC.
• MNSeg's architecture is adaptable and generalizes well with various 3D segmentation backbones.