Self-Attention Similarity Guided Graph Convolutional Network for Multi-type Lower-Grade Glioma Classification Research

Neurology Computer Science Information Science Artificial Intelligence Medicine Oncology Medical Imaging
low-grade glioma brain tumor classification 3D brain MRI graph convolutional network self-attention

Self-Attention Similarity Guided Graph Convolutional Network

Graph Convolutional Network Based on Self-Attention Similarity for Multi-type Low-Grade Glioma Classification

1. Research Background

Low-grade glioma is a common malignant brain tumor caused by the cancerous transformation of glial cells in the brain and spinal cord. Gliomas are characterized by high incidence, high recurrence rate, high mortality rate, and low cure rate. Correct classification of multi-type low-grade gliomas is crucial for patient prognosis. In diagnosis, doctors typically use Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) to analyze the Isocitrate Dehydrogenase (IDH) mutation status of glioma cells.

The IDH mutation status is an important marker for distinguishing wild-type and mutant gliomas. Traditionally, it needs to be determined through biopsy or surgical resection, followed by immunohistochemistry or gene sequencing. Since biopsies carry certain risks, developing a non-invasive method for predicting IDH mutation status with computer-aided diagnosis is of significant importance, avoiding unnecessary surgical risks for patients.

2. Paper Source

This paper was published in the IEEE Journal of Biomedical and Health Informatics in the July 2023 issue. The authors are from the School of Information and Management Science, Henan Agricultural University.

3. Research Content

1) Research Process

This study proposes a Self-Attention Similarity Guided Graph Convolutional Network (SASG-GCN) for classifying tumor foci (TF), mutant gliomas (MG), and wild-type gliomas (WTG) from MRI images. The workflow of SASG-GCN includes:

  1. Using Convolutional Deep Belief Network (CDBN) to extract feature vectors from sequential MRI slices, serving as nodes to construct the graph network.

  2. Proposing a self-attention similarity (SAS) method to calculate similarity coefficients between feature vectors, acting as edge weights in the graph network.

  3. Constructing a graph network with 64 nodes and 835 edges based on the extracted feature vectors and similarity coefficients.

  4. Inputting the constructed graph network into a two-layer Graph Convolutional Network (GCN) for classification prediction.

2) Main Results

The research was experimentally evaluated on the TCGA-LGG dataset (including 402 MRI images), and the results indicate that:

  • SASG-GCN achieved the highest accuracy of 93.62% in the tri-classification task, outperforming various other state-of-the-art methods.
  • Compared to Convolutional Neural Networks based on 2D slices or 3D images, SASG-GCN better captures the nonlinear relationships and high-dimensional information in MRI images.
  • The self-attention similarity guided strategy significantly enhanced classification performance.
  • Experimental visualizations displayed distinctive features of different types of gliomas in the constructed graph network.

3) Research Significance

The main innovations and significance of this study include:

  • Proposing a novel graph convolutional network-based method for multi-class glioma classification.
  • Utilizing CDBN to reduce the dimensionality of high-dimensional MRI image data and SAS to capture the similarity relationships between slices, guiding the construction of a discriminative graph network representation together.
  • Experiments demonstrated the superior performance of SASG-GCN in glioma classification tasks, providing new insights into MRI image processing.
  • Visual analysis revealed significant structural differences in the graph network for different types of gliomas, offering guidance for clinical diagnosis.

4) Research Highlights

  • Proposing an innovative glioma classification framework based on graph convolutional networks.
  • Effectively extracting high-dimensional data features and inter-slice associations using a self-attention similarity guided strategy.
  • Experiments showed that the proposed method achieved the latest optimal performance in multi-class glioma classification tasks.
  • The construction of graph networks and visual analysis provided new interpretative methods for clinical diagnosis.

4. Conclusion

This study proposes an innovative SASG-GCN model that achieves non-invasive precise classification of low-grade gliomas by constructing a discriminative graph network representation encompassing slice semantics and inter-slice similarity information. The research outcomes not only theoretically expand the application of graph convolutional networks in medical image processing but also provide new computer-aided tools for clinical glioma diagnosis, holding significant scientific value and application potential.