CLAIRE-Parallelized Diffeomorphic Image Registration for Large-Scale Biomedical Imaging Applications.

Area of Science:

• Biomedical Imaging
• Medical Image Analysis
• Computational Anatomy

Background:

• Diffeomorphic image registration is crucial for biomedical applications.
• Existing software struggles with large-scale, high-resolution biomedical images due to computational cost.
• Downsampling is a common workaround, but its impact on registration accuracy is not fully understood.

Purpose of the Study:

• To evaluate the performance of the CLAIRE algorithm for large-scale, high-resolution image registration.
• To analyze the impact of downsampling on registration accuracy in biomedical imaging.
• To compare full-resolution registration with downsampled registration using synthetic and real-world datasets.

Main Methods:

• Utilized CLAIRE, a multi-node, multi-GPU diffeomorphic image registration algorithm.
• Performed registrations at full resolution and various downsampled resolutions (e.g., 1024^3 to 256^3).
• Compared registration quality using metrics like the Dice coefficient on synthetic and real-world biomedical imaging datasets.

Main Results:

• Full-resolution registration with CLAIRE can yield superior registration quality compared to downsampled approaches.
• Downsampling significantly decreased registration accuracy (e.g., Dice coefficient from 92% to 79% for a synthetic image).
• The impact of downsampling was less pronounced on noisy or low-contrast high-resolution images.

Conclusions:

• CLAIRE enables registration of clinically relevant sized images in seconds and unprecedented resolutions in reasonable time.
• Registration at full resolution is not always superior, depending on image properties.
• This study provides the first analysis of image registration quality at resolutions up to 2816×3016×1162.