Abstract
BACKGROUND: Metallic interventional devices such as brachytherapy seeds and stents, are extensively utilized in clinical settings. However, these devices generate significant susceptibility artifacts in conventional magnetic resonance imaging (MRI), manifesting as signal voids that impede precise visualization. Susceptibility-driven positive contrast MRI (PC-MRI) mitigates this limitation by solving regularized ℓ1-norm minimization problems to reconstruct positive contrast images. The conventional nonlinear conjugate gradient (CG) algorithm, commonly employed for solving such non-smooth convex optimization problems, encounters challenges, including slow convergence rates, sensitivity to initial solutions and parameter selection, and difficulties in achieving optimal imaging reconstruction due to ill-posed inversion problems. This study aimed to develop and evaluate an accelerated primal-dual (PD) optimization framework with graphics processing unit (GPU) parallelization to overcome the limitations of the conventional CG algorithm for susceptibility-driven PC-MRI reconstruction. The proposed method seeks to solve the exact ℓ1-minimization problem without smoothing approximations. METHODS: The efficacy of the method was evaluated through computational simulations, phantom experiments, and in vivo studies. Quantitative assessments included convergence behavior, full width at half maximum (FWHM), signal-to-noise ratio (SNR), and reconstruction time. Statistical significance was determined using paired t-tests, with a significance threshold set at P<0.01. RESULTS: Comparing to the conventional CG method, the PD approach can provide a faster reconstruction convergence rate of 2-4 times, and it demonstrated an end-to-end easy-adjustment method that does not rely on parameter tuning. The results also show that the PD method achieves better visualization and more accurate localization of the metallic interventional devices in positive contrast. Quantitative evaluations showed that the PD method achieved a significant reduction in FWHM near metallic seeds (e.g., from 1.15 for CG to 1.02 for PD in Patient 2, 11.3% improvement, P<0.01), indicating superior image sharpness. A notable improvement in SNR was also observed (e.g., from 110.34 for CG to 131.65 for PD in Patient 2, 19.3% enhancement, P<0.01), confirming enhanced image quality in both phantom and in vivo experiments. Furthermore, GPU acceleration further improved reconstruction speed of the PD approach by 4-15 times. CONCLUSIONS: The susceptibility-driven positive contrast imaging technique based on PD regularization demonstrates faster convergence, superior image quality, and easier parameter adjustment compared to conventional CG methods. The speed of reconstruction can be further improved by GPU acceleration.