Full domain-decomposition scheme for diffuse optical tomography of large-sized tissues with a combined CPU and GPU...

Area of Science:

• Biomedical Optics
• Computational Imaging
• Medical Physics

Background:

• Diffuse optical tomography (DOT) commonly solves nonlinear, ill-posed inverse problems.
• Linearized iterative methods in DOT face computational and storage challenges for large tissues.
• Current methods limit the use of matrix-based linear inversions for better image quality.

Purpose of the Study:

• To develop a computationally efficient and scalable parallelized scheme for diffuse optical tomography.
• To overcome the limitations of traditional iterative approaches in DOT for large-scale applications.
• To improve the quantitative performance of DOT reconstructions.

Main Methods:

• A parallelized full domain-decomposition scheme is proposed, dividing the domain into overlapped subdomains.
• Sub-inversions are solved independently using Schwarz-type iterations.
• A combined multicore CPU and multithread graphics processing unit (GPU) parallelization strategy is employed.

Main Results:

• The proposed method effectively reduces computation time for large-sized DOT problems.
• Significant reduction in memory occupation was observed.
• Improved quantitative performance in image reconstruction was demonstrated through numerical and phantom experiments.

Conclusions:

• The parallelized domain decomposition scheme offers a viable solution for computationally intensive DOT applications.
• This approach enhances the feasibility of DOT for large tissue imaging, such as in breast tumor diagnosis and brain functional imaging.
• The method improves image reconstruction quality and efficiency, paving the way for broader clinical and research use.