Spatiotemporal Gaussian representation-based dynamic reconstruction and motion estimation framework for time-resolved...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Time-resolved volumetric MRI is crucial for motion-adaptive radiotherapy to track anatomical changes during treatment.
• Existing methods often require prior anatomical or motion models, limiting their applicability.

Purpose of the Study:

• To develop a novel framework, DREME-GSMR, for time-resolved dynamic MRI reconstruction without prior models.
• To enable real-time intra-treatment volumetric MR imaging and motion tracking for improved radiotherapy accuracy.

Main Methods:

• Developed a spatiotemporal Gaussian representation framework (DREME-GSMR) using 3D Gaussians for anatomy and motion.
• Incorporated a dual-path MLP/CNN motion encoder to estimate motion coefficients from k-space data.
• Introduced a motion-augmentation strategy for enhanced robustness during real-time imaging.

Main Results:

• DREME-GSMR achieved ~400ms temporal resolution with ~10ms/volume inference time.
• Demonstrated high accuracy in phantom and clinical datasets, with mean liver center-of-mass error as low as 0.96 mm in patients.
• Showcased strong performance in dynamic reconstruction and real-time imaging, with SSIM scores up to 0.92.

Conclusions:

• DREME-GSMR enables accurate, real-time dynamic MRI reconstruction for motion-adaptive radiotherapy.
• The framework's ability to work without prior models and its robustness offer significant advantages for clinical translation.
• This technology has the potential to enhance precision and efficacy in radiotherapy by adapting to patient motion.