Accelerated magnetic resonance imaging reconstruction using mixed-precision and specialized hardware
Active Project Grant Awarded Last action
- Award ID
- F32EB040322
- Recipient
-
The Regents Of The University Of California. San Francisco UEI KMH5K9V7S518
- Awarding agency
-
National Institute of Biomedical Imaging and Bioenginering
National Institutes of Health, Department of Health and Human Services
- Obligated
- $79,000
Award facts
- Assistance listing
- 93.286 Discovery and Applied Research for Technological Innovations to Improve Human Health
- Completion date 0% complete
- Place of performance
- San Francisco, CA 94143, USA
Award details come from USAspending.gov.
Description
- Abstract
- Advanced MRI reconstruction algorithms including parallel imaging, compressed sensing, and deep learning methods offer transformative improvements in image quality and scan acceleration. However, clinical translation of these techniques is hindered by lengthy reconstruction times, high computational demands, and hardware constraints, particularly for ultra-high-field imaging, real-time MR-guided interventions, and portable MRI systems. Current reconstruction pipelines rely on uniform 32-bit floating-point (FP32) precision throughout all operations, representing a missed opportunity for computational optimization. The long-term goal of this research is to create computation- and power-efficient reconstruction platforms that enable widespread clinical adoption of advanced MRI techniques across diverse clinical settings. The objective of this fellowship is to design, implement, and validate a hardware/software CO-design framework for mixed- precision MRI reconstruction. The central hypothesis is that systematic replacement of FP32 operations with low-precision FP8/FP16 formats in both software and custom hardware will achieve up to 4× speedup in reconstruction throughput and up to 10× reduction in latency compared to FP32 GPU baselines, without sacrificing diagnostic image quality. Aim 1 will develop and validate mixed-precision algorithms for accelerated MRI reconstruction on next-generation gpus. This includes implementing mixed-precision coil sensitivity estimators and operators, designing a novel microscaling (MX) blocked FFT algorithm to leverage FP8 tensor cores, and evaluating diagnostic quality and memory efficiency across publicly available datasets, ultra-high field datasets, and portable MRI datasets. Aim 2 will design and prototype fpga-based hardware CO-processors for low-latency MRI reconstruction. This involves implementing deep learning inference on a risc-v system-on-chip using custom machine learning accelerators and integrating custom FP8-accelerated FFT modules to create a complete parallel CO-processor for unrolled reconstruction networks, and evaluating computation speed, memory access patterns, power usage, and latency compared to CPU/GPU baselines. Health relevance: this research directly addresses nibib's mission to improve human health by developing enabling technologies for medical imaging. Mixed-precision algorithms and custom hardware accelerators will democratize access to advanced MRI reconstruction across clinical settings, from high-end academic centers with ultra-high field scanners to resource-limited environments using portable MRI systems. By enabling sophisticated reconstruction on modest hardware, this work will expand MRI access globally and improve patient outcomes.
Quick answers
Who received award F32EB040322?
F32EB040322 was awarded to The Regents Of The University Of California. San Francisco by National Institute of Biomedical Imaging and Bioenginering (National Institutes of Health) on February 1, 2027.
How much is award F32EB040322 worth?
$79,000 has been obligated on F32EB040322.
On GovTribe
See this award's full record
- Every transaction and subaward
- The recipient's other awards and teaming partners
- Similar grant awards
- Ask GovTribe AI about this award