Image reconstruction by Mumford-Shah regularization for low-dose CT with multi-GPU acceleration

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Mumford-Shah (MS) functional is a regularization technique for X-ray computed tomography (CT).
• High-resolution CT applications face implementation challenges due to large data sizes.
• Existing methods struggle with the computational demands of large-scale 3D reconstructions.

Purpose of the Study:

• To implement and accelerate MS regularization on a multi-GPU platform for large-scale 3D CT reconstruction.
• To address the limitations of data size and computational complexity in high-resolution CT.
• To improve reconstruction quality and efficiency for ultra-low-dose CT imaging.

Main Methods:

• Developed a novel 3D volume partition scheme for multi-GPU parallel processing.
• Implemented a multithread parallel acceleration strategy to utilize multi-GPU resources.
• Designed a modular and scalable implementation adaptable to various GPU configurations.

Main Results:

• Achieved superior reconstruction quality compared to total variation regularization, particularly in ultra-low-dose scenarios.
• Successfully reconstructed huge 3D images (up to 3072^3 voxels) within 12 minutes.
• Demonstrated efficient utilization of multi-GPU computing resources for large-scale image reconstruction.

Conclusions:

• The proposed multi-GPU accelerated MS regularization effectively resolves implementation difficulties for large-scale 3D CT.
• This method offers significant advantages in reconstruction quality and speed, especially for ultra-low-dose CT.
• It represents the first successful application of MS regularization for reconstructing extremely large 3D CT images within a practical timeframe.