Adaptive Riemannian optimization for multi-scale diffeomorphic matching

Area of Science:

• Biomedical imaging
• Computational biology
• Medical image analysis

Background:

• Accurate image matching is crucial for analyzing biomedical and biological data.
• Current registration methods are slow and deep learning approaches require extensive training and memory.

Purpose of the Study:

• To develop a fast, accurate, and training-free image matching algorithm.
• To address the limitations of existing registration techniques.

Main Methods:

• Proposed FireANTs, a GPU-accelerated, multi-scale adaptive Riemannian optimization algorithm.
• Implemented a training-free approach for dense diffeomorphic image matching.

Main Results:

• FireANTs doubles registration speed on CPU and is 100x faster on GPU compared to ANTs.
• Achieved competitive inference runtime with deep learning methods on GPU, using 10x less memory.
• Demonstrated robustness across diverse modalities, species, and organs without tuning.

Conclusions:

• FireANTs provides a significant advancement in image registration speed and accuracy.
• The training-free, GPU-accelerated framework reduces computational resources for research and development.