Learning discrete structures for cancer radiomics.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Radiomics methods extract quantitative imaging features for cancer analysis.
• Current methods often ignore underlying image relationships, limiting performance.
• Existing graph neural networks (GNNs) struggle with unknown, task-specific image relations.

Purpose of the Study:

• To develop a novel Image-Graph based neural network (IGNN) for cancer image analysis.
• To learn image relationships and refine features simultaneously by minimizing task-specific loss.
• To address scenarios with unknown image relations and the need for task-specific graph learning.

Main Methods:

• Developed an Image-Graph based neural network (IGNN).
• Learned discrete image graph structures and refined features concurrently.
• Utilized task-specific loss minimization for graph and feature learning.

Main Results:

• Achieved superior area under the curve compared to existing radiomics and GNNs across four real datasets.
• Demonstrated effective learning of task-specific graphs for different datasets.
• Validated performance across data from five different hospitals.

Conclusions:

• The proposed Image-Graph based neural network (IGNN) effectively captures unknown, task-specific image relationships.
• IGNN offers improved performance in cancer image analysis over current state-of-the-art methods.
• This approach enhances quantitative imaging feature extraction by leveraging learned image-graph structures.