PhysMorph: A biomechanical and image-guided deep learning framework for real-time multi-modal liver image...

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Accurate registration of pretreatment Magnetic Resonance Imaging (MRI) to onboard Cone Beam Computed Tomography (CBCT) is crucial for liver Stereotactic Body Radiation Therapy (SBRT).
• Challenges include poor CBCT soft-tissue contrast and respiratory motion, impacting treatment precision.
• Existing methods may lack the speed and accuracy required for real-time clinical application.

Purpose of the Study:

• To develop and validate PhysMorph, a physics-informed deep learning framework for rapid and anatomically plausible MR-CBCT image registration of the liver.
• To improve the accuracy and efficiency of image registration for liver SBRT.
• To enable real-time application of advanced image registration techniques in radiotherapy.

Main Methods:

• Developed PhysMorph, a framework integrating finite element method (FEM) simulations for biomechanical regularization with image similarity metrics.
• Validated the framework on simulated data with known ground-truth deformation and clinical MR-CBCT pairs from liver SBRT patients.
• Assessed performance using target registration error (TRE), mean surface distance (MSD), and biomechanical fidelity metrics.

Main Results:

• PhysMorph achieved a mean TRE of 2.2 ± 1.4 mm and MSD of 1.60 ± 0.05 mm on clinical data.
• Significantly outperformed VoxelMorph and SynthMorph in registration accuracy while maintaining high biomechanical fidelity.
• Reduced registration time to 103.4 ms, enabling practical real-time application, a substantial improvement over conventional FEM methods.

Conclusions:

• PhysMorph provides fast, accurate, and physically realistic registration of pretreatment MRI to on-board CBCT for liver SBRT.
• The framework's integration of MRI's soft-tissue visualization and anatomical plausibility facilitates precise tumor localization.
• This advancement may allow for smaller planning target volumes, more conformal dose distributions, enhanced tumor control, and reduced radiation exposure to healthy tissues.