An implicit neural deformable ray model for limited and sparse view-based spatiotemporal reconstruction.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Radiation Oncology

Background:

• Continuous spatiotemporal volumetric reconstruction is crucial for radiation therapy to analyze treatment outcomes.
• Current methods face limitations due to radiation exposure and hardware, often relying on intermittent imaging.
• Accurate tracking of patient anatomy during treatment is essential for dose calculation and outcome prediction.

Purpose of the Study:

• To develop a method for spatiotemporal reconstruction from limited X-ray views.
• To achieve continuous-time dynamic volume reconstruction using minimal input projections.
• To enable patient-specific volumetric reconstruction with as few as 20 views and real-time tracking with two orthogonal views.

Main Methods:

• Introduced the implicit neural deformable ray (INDeR) model.
• Utilized a ray bundle coordinate system and embedded sparse view measurements into an implicit neural field.
• Estimated real-time motion via low-dimensional modulation for deformable ray bundles.

Main Results:

• The INDeR model achieved robust image reconstruction and motion tracking.
• Demonstrated high peak signal-to-noise ratio (PSNR) of 30.13 dB with 20 views, outperforming existing methods.
• Maintained a PSNR > 27.41 dB in real-time using only two orthogonal projections.

Conclusions:

• The INDeR framework successfully reconstructs continuous spatiotemporal representations from sparse views.
• Achieved highly accurate reconstruction with minimal projections and effective real-time tracking.
• Shows significant potential for anatomical monitoring in radiation therapy.