3D nonrigid registration via optimal mass transport on the GPU

Area of Science:

• Medical image analysis
• Computational mathematics
• Computer vision

Background:

• Optimal mass transport (OMT) is crucial for image registration but computationally intensive.
• Previous OMT methods for non-rigid registration face challenges with speed and accuracy.
• Accurate 3D image registration is vital for medical applications like surgical planning.

Purpose of the Study:

• To develop a computationally efficient numerical scheme for OMT.
• To apply this scheme to non-rigid 3D image registration.
• To improve the accuracy and speed of medical image registration.

Main Methods:

• A novel minimizing flow approach for OMT.
• Utilization of all grayscale image data, avoiding landmark specification.
• Implementation using multigrid and parallel processing on Graphics Processing Units (GPUs).

Main Results:

• The new method is orders of magnitude faster than previous OMT techniques.
• Achieved previously unattainable optimality measures (mean curl) for transport maps.
• Demonstrated accurate non-rigid registration of 3D synthetic and real brain MRI data.

Conclusions:

• The proposed OMT scheme offers a significant advancement in computational efficiency for 3D image registration.
• The method enhances registration accuracy, particularly for complex non-rigid transformations.
• This approach has strong potential for clinical applications in medical image analysis.