Compressed sensing (CS) is a signal processing technique that enables faster scan times in medical imaging. Philips Healthcare integrated CS methods into their magnetic resonance imaging (MR) scanners to reduce scan time by up to 50 percent for 2D and 3D sequences, compared to Philips scans without Compressed SENSE, with virtually equal image quality. Recently, deep learning methods have been explored for reconstructing MRI images, showing good results in terms of image quality and speed of reconstruction. Philips Healthcare and Intel report on two hybrid frequency-domain/image-domain encoder/decoder architectures that produce excellent results in MRI reconstruction.
We show how these two neural networks can be accelerated on Intel® hardware through use of the Intel® Distribution of OpenVINO™ Toolkit. The toolkit allows Philips Healthcare to speed up their deep learning inference by as much as 54x over standard, unoptimized TensorFlow 1.15, as tested in Philips’ proprietary Linux environment on Intel® Xeon® processors.1 We further describe how to leverage the Intel® DevCloud for the Edge, which allowed Philips Healthcare to compare performance of their deep learning models on Intel® Xeon® and Intel® Core™ processors, Intel® Movidius™ Vision Processing Units (VPUs), FPGAs, and integrated GPU hardware in order to design deep learning products of various performance, price, power, and form factors.
Read the white paper: Philips Healthcare Uses the Intel® Distribution of OpenVINO™ Toolkit and the Intel® DevCloud for the Edge to Accelerate Compressed Sensing Image Reconstruction Algorithms for MRI.