Accelerated nonrigid intensity-based image registration using importance sampling

Area of Science:

• Medical image analysis
• Computational anatomy
• Scientific computing

Background:

• Nonrigid image registration is crucial for estimating deformations in medical imaging.
• Intensity-based similarity metrics and gradient optimization are common but computationally intensive for large datasets.
• Stochastic gradient approximation using random voxel subsets can reduce computation time.

Purpose of the Study:

• To introduce an importance sampling framework to reduce the variance of stochastic gradient approximations in nonrigid image registration.
• To develop an edge-dependent adaptive sampling distribution for intensity-based registration algorithms.
• To evaluate the efficiency and accuracy of importance sampling in accelerating registration.

Main Methods:

• Implemented an importance sampling strategy for gradient approximation in nonrigid registration.
• Utilized an edge-dependent adaptive sampling distribution tailored for intensity-based metrics.
• Compared stochastic approximation with and without importance sampling against deterministic gradient descent.
• Tested on simulated MRI brain data and real CT lung data from eight subjects.

Main Results:

• Importance sampling significantly reduced the variance of gradient approximations.
• The combination of stochastic approximation and importance sampling accelerated the registration process.
• Accuracy of registration was preserved compared to deterministic methods.
• Demonstrated effectiveness on both simulated and real medical imaging datasets.

Conclusions:

• Importance sampling is an effective technique for accelerating nonrigid image registration.
• The proposed edge-dependent adaptive sampling improves efficiency without compromising accuracy.
• This approach offers a viable solution for computationally demanding registration tasks in medical imaging.