A unified generation-registration framework for improved MR-based CT synthesis in proton therapy

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Computational Anatomy

Background:

• Magnetic Resonance (MR) imaging aids proton therapy planning by generating synthetic Computed Tomography (sCT) images.
• Accurate MR-CT image alignment is crucial but challenging, especially in mobile areas like the head-and-neck.
• Misalignment leads to blurred sCTs, compromising treatment precision and effectiveness.

Purpose of the Study:

• To develop a novel network unifying image generation and registration for enhanced sCT quality and anatomical fidelity.
• To improve the alignment of MR and CT images for more accurate proton therapy planning.

Main Methods:

• A unified network synergizes a UNet-based generation network (G) and an implicit neural representation (INR) deformable registration network (R).
• Joint optimization alternately minimizes discrepancies between generated/registered CTs and reference CTs.
• Validation performed on 60 head-and-neck patient datasets, with 12 reserved for testing.

Main Results:

• The proposed unified network achieved a Mean Absolute Error (MAE) of 80.98 ± 7.55 HU, significantly outperforming the baseline Pix2Pix (124.95 ± 30.74 HU).
• Generated sCTs exhibited sharper anatomical details and improved congruence compared to baseline methods.
• Recalculated proton therapy plans on the generated sCTs showed reduced dosimetric discrepancies compared to reference plans.

Conclusions:

• A holistic, unified approach integrating MR-to-CT synthesis and deformable registration significantly enhances sCT precision and quality.
• This method is particularly effective for challenging anatomical regions with significant MR-CT variations.
• The improved sCTs facilitate more accurate proton therapy treatment planning.