Graph based multi-scale neighboring topology deep learning for kidney and tumor segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate spatial and semantic relation modeling is crucial for volumetric image segmentation.
• Existing methods face challenges in capturing multi-range dependencies between image regions.

Purpose of the Study:

• To develop a novel deep graph reasoning model for enhanced volumetric image segmentation.
• To effectively learn spatial dependencies and semantic connections across diverse image regions.

Main Methods:

• Constructing a graph with image regions as nodes and deriving topology for spatial and semantic connections.
• Employing multi-order random walks for node attribute embedding and multi-scale graph convolutional autoencoders for representation extraction.
• Integrating a scale-level attention module for adaptive fusion of topological representations.

Main Results:

• The proposed model significantly outperforms state-of-the-art methods on public kidney and tumor CT segmentation datasets.
• Ablation studies confirm the effectiveness of the model's key innovations and its generalization ability across different backbones.

Conclusions:

• The novel random walk-driven multi-scale graph convolutional network with scale-level attention accurately extracts semantic connections and spatial dependencies.
• This approach enables precise kidney and tumor segmentation in CT volumes.