Faster dense deformable image registration by utilizing both CPU and GPU.

Area of Science:

• Medical image analysis
• Computational imaging
• High-performance computing

Background:

• Deformable image registration is crucial for medical image analysis tasks like image fusion and segmentation.
• Large-scale datasets necessitate efficient registration methods due to high computational costs.
• Existing methods like Advanced Normalization Tools (ANTs) can be computationally intensive.

Purpose of the Study:

• To accelerate a deformable image registration method using a heterogeneous computing approach (CPU and GPU).
• To improve the efficiency and labeling quality of medical image registration for large-scale databases.

Main Methods:

• Implemented a hybrid CPU-GPU strategy, offloading matching criterion computation to the GPU and optimization to the CPU.
• Utilized a pipeline model to overlap computational tasks and minimize data synchronization overhead.
• Evaluated performance on a brain labeling task, comparing against a CPU-only implementation and ANTs software.

Main Results:

• Achieved speed-up factors of 4x against the CPU implementation and 8x against ANTs.
• Observed improved labeling quality with mean Dice overlaps of 0.712 for the proposed method versus 0.701 for ANTs.
• Demonstrated favorable comparison to ANTs in both speed and accuracy.

Conclusions:

• The proposed heterogeneous computing approach significantly accelerates deformable image registration.
• The method offers improved labeling quality compared to established software like ANTs.
• The registration method and parallelization strategy are released as open-source software (deform).