سامانه همایش ها

صفحه اصلی / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

GPU-Accelerated GRAPPA: A Fast Implementation Using PyTorch for MRI Reconstruction

نویسندگان :

Mehrdad Anvari-Fard¹ Mahdi Bazargani² Mohammad Javad Heidari³ Hamid Soltanian-Zadeh⁴

1- School of Electrical and Computer Engineering, College of Engineering, University of Tehran Tehran, Iran 2- School of Electrical and Computer Engineering, College of Engineering, University of Tehran Tehran, Iran 3- School of Electrical and Computer Engineering, College of Engineering, University of Tehran Tehran, Iran 4- School of Electrical and Computer Engineering, College of Engineering, University of Tehran Tehran, Iran

کلمات کلیدی :

GRAPPA،MRI Reconstruction،Deep Learning،FastMRI،GPU acceleration

چکیده :

GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA) is a widely used algorithm in MRI parallel imaging that reconstructs accelerated MRI scans by estimating the unknown phase-encoding lines omitted during k-space data acquisition. Unlike SENSE (Sensitivity Encoding), which operates in the image domain, GRAPPA directly processes k-space data and offers high reconstruction quality without requiring prior knowledge of coil sensitivity maps, making it one of the most commonly used algorithms for MRI reconstruction in clinical practice. Recent MRI reconstruction trends increasingly combine classical methods with deep learning, either as end-to-end trainable networks or hybrid pipelines that use physics-based operators within learning frameworks. GRAPPA is often employed as a preprocessing step before feeding slice information into deep learning models for MRI reconstruction. Despite its effectiveness, GRAPPA is typically a time-consuming part of the training process. In this work, we leverage the GPU capabilities of the PyTorch library and employ several optimization techniques to accelerate the GRAPPA algorithm. Our implementation is compared against the PyGRAPPA repository, developed by Nicholas McKibben, using a subset of the NYU fastMRI dataset. The results demonstrate that our optimized implementation achieves more than 40-fold speedup, which is statistically significant (p < 0.01) while maintaining equivalent image quality with no significant differences in reconstruction metrics (p > 0.05).