Known Operator Learning Enables Constrained Projection Geometry Conversion: Parallel to Cone-Beam for Hybrid MR/X-Ray...

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Simultaneous X-ray and Magnetic Resonance Imaging (MRI) offer enhanced guidance for interventional procedures like stroke therapy.
• A key challenge is reconciling the differing acquisition geometries: MRI's parallel projection and X-ray's perspective projection.
• Conventional rebinning methods for this conversion lead to significant resolution loss.

Purpose of the Study:

• To develop a novel rebinning algorithm for accurate parallel-to-cone-beam conversion.
• To leverage deep neural networks for learning unknown operators in the conversion process.
• To evaluate the performance and generalizability of the proposed deep learning approach.

Main Methods:

• A novel rebinning formula was derived for parallel-to-cone-beam conversion.
• A deep neural network architecture was designed based on the operator learning paradigm with differentiable projection operators.
• The network was trained using simulated data, and its performance was compared against classical rebinning methods.

Main Results:

• The developed deep neural network demonstrated superior performance compared to the classical rebinning approach.
• The study confirmed that the network can be trained effectively with simulated data without compromising generality.
• The trained operators are applicable to real-world data without requiring retraining or transfer learning.

Conclusions:

• The proposed deep learning-based rebinning algorithm effectively addresses the resolution loss issue in MRI-to-X-ray conversion.
• The method shows promise for improving real-time image guidance in interventional procedures.
• Data-driven learning with simulated data provides a robust and generalizable solution for complex image reconstruction tasks.