Hardware acceleration for fast Magnetic Resonance Fingerprinting map reconstruction: FPGA porting of a deep learning algorithm

About

Magnetic Resonance Fingerprinting (MRF) is a fast quantitative MR Imaging technique able to obtain multi-parametric maps with a single acquisition, but data processing is limited by escalating memory and computation needs. Neural Networks (NNs) accelerate reconstruction, but training still requires significant resources. We propose an FPGA-based NN for real-time brain parameter reconstruction from MRF data. After a traditional software validation, the NN is reduced through Quantization Aware Training to meet the available resources of the FPGA hardware accelerator, creating a quantized model that uses lower precision without affecting the NN performance. Training the NN is estimated to take 1000 to 10000 seconds, representing a significant improvement over standard CPU-based training, which can be up to 36 times slower. This approach has the potential to enable real-time brain analysis on mobile devices, potentially revolutionizing clinical decision-making and telemedicine..

0:00:00 Slide 1

0:01:55 Slide 2

0:03:01 Slide 3

0:05:51 Slide 4

0:08:30 Slide 5

0:09:53 Slide 6

0:11:09 Slide 7

0:13:58 Slide 8