Accelerating iterative coordinate descent using a stored system matrix.

Area of Science:

• Medical Imaging
• Computational Science
• Image Reconstruction

Background:

• Model-based iterative reconstruction (MBIR) faces computational challenges.
• Iterative coordinate descent (ICD) has been perceived as incompatible with modern hardware like GPUs.

Purpose of the Study:

• To accelerate the open-source FreeCT_ICD algorithm by incorporating GPU acceleration.
• To demonstrate that GPU-accelerated ICD can achieve computational performance competitive with simultaneous update approaches for MBIR.

Main Methods:

• Implemented GPU acceleration for FreeCT_ICD by optimizing sinogram memory access and enabling simultaneous voxel updates (N_S > 1).
• Empirically studied the convergence behavior of ICD with simultaneous updates on a clinical dataset, as theoretical guarantees are lost.

Main Results:

• GPU-accelerated ICD achieved reconstruction times of ~20 seconds per iteration on a single GPU, compared to 2300 seconds per iteration on a 6-core CPU.
• The algorithm converged within 2 HU RMS difference after 400 iterations, and within 10 HU RMSD in 4 minutes with wFBP initialization.
• Maximum performance was achieved with N_S = 16, with no observed divergence until N_S > 1024.

Conclusions:

• Iterative coordinate descent (ICD), with appropriate GPU optimization, can achieve computational performance competitive with existing simultaneous update algorithms for MBIR.
• This work demonstrates the practical viability of ICD on modern computing architectures for accelerated image reconstruction.