Dense motion propagation from sparse samples.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Biophysics

Background:

• Sparse sampling of dynamic medical imaging data is crucial for motion estimation.
• Existing methods face challenges in achieving high temporal and spatial resolution for dynamic processes like respiratory motion.

Purpose of the Study:

• To develop a generalized framework for dense motion propagation from sparse samples using transfer learning and manifold alignment.
• To animate high-resolution static MR images with 4D free-breathing respiratory motion derived from low-resolution sparse data.
• To demonstrate the viability of sparse sampling for building high-resolution models of free-breathing respiratory motion.

Main Methods:

• A generalized framework for dense motion propagation from sparse samples.
• Utilizing transfer learning and manifold alignment to transfer knowledge across different imaging domains.
• Employing a propagation model in the eigen domain to estimate 4D motion vector fields from sparsely sampled dynamic MRI data.
• Constrained articulation to build free-breathing respiratory motion models.

Main Results:

• Demonstrated that sparse sampling of dynamic MRI can successfully build models of free-breathing respiratory motion.
• Achieved high contrast, temporal, and spatial resolution in articulated motion models, surpassing conventional methods.
• Estimated motion vector fields from sparse sampling showed equivalence to fully sampled 4D dynamic data with an RMS error of approximately 2 mm.
• Presented exemplar 4D high-contrast, high-resolution articulated volunteer datasets.

Conclusions:

• Sparse sampling of free-breathing data within a manifold alignment and transfer learning paradigm can estimate fully sampled motion.
• The proposed approach offers greater acquisition freedom for free-breathing respiratory motion sequences.
• The methodology can be applied to other sparse sampling scenarios for dense motion propagation.