Faster 4D Imaging with Heterogenous Computing

Medical imaging is one of the most valuable tools doctors have in detecting and diagnosing disease or anomalies in patients. Both 2D and 3D imaging play a big role in enabling medical professionals to deliver better clinical outcomes, from ultrasound that provides fast 2D images to computed tomography (CT) and magnetic resonance imaging (MRI) that provide highly accurate 3D images of the human body. The emergence of 4D imaging, however, is taking medical imaging into the next frontier with moving pictures. For example, 4D imaging is being used in respiratory analysis where an MRI can analyze respiration cycles. Such advances in medical imaging are exciting, but do not come without challenges. 4D MRI imaging requires significant pre- and post-processing to reconstitute an image.

An MRI scan consists of two elements: the scan during to acquire data, followed by reconstruction. During the scan, the data samples are captured along a predefined trajectory. These samples are spatial in nature and are in what is called the k-space domain. Transforming the acquired samples into an understandable image occurs in the reconstruction phase. MRIs, therefore, face the competing difficulties of generating high-definition imaging, low signal-to-noise and fast scan time.

The complexity of the image reconstruction depends upon the sampling trajectory. A simple Cartesian scan direction provides the k-space samples aligned to a grid, allowing quick image reconstruction using a Fast Fourier Transform. A non-Cartesian scan, like a spiral trajectory for example, will result in the k-space samples aligned in a more complex pattern which requires advanced image reconstruction algorithms. Currently this can take several minutes for the image to be available after the scan is completed and requires considerable processing capabilities. This makes deploying 4D imaging solutions difficult, hindering wider adoption. For research purposes, server farms can be used to demonstrate the algorithm performance. However, a deployable solution needs a computational capability that can perform both the RF signal drive, signal capture and image reconstruction.

One solution enabling wider adoption of 4D imaging are programmable SoCs. Xilinx heterogeneous Zynq UltraScale+ MPSoCs help to address the simultaneous challenges facing 4D imaging through the advantages of an integrated high-performance processor system with parallel programmable logic.

Thanks to the unique architecture of these devices, the programmable logic can be used to interface with both the RF drive waveform and ADCs to capture the resultant data from the scan. This wide-ranging interfacing capability also enables parallel data structures to be implemented within the programmable logic to support multiple parallel high-speed RF generation or signal capture. At the same time, the processing system can be used to generate user interfaces, communicate with medical records systems and more.

Both RF drive waveforms and image reconstitution can be accelerated using programmable logic resources. As these algorithms are complex, developers can improve productivity by using Xilinx Vitis, a high-level synthesis (HLS) tool which allows engineers to develop algorithms in C/C++ or OpenCL without the need to work at the hardware description level. Using Vitis HLS the developer can define the algorithms at a higher level and exploit the parallel nature of programmable logic for example unrolling loops, and pipelining operations to take advantage of the parallelism which exists in the algorithms. Implementing the algorithms in programmable logic can provide a significant performance boost.

Summary

Medical imaging is a key technology that has enabled medical professionals to understand in greater detail the human body and advance treatments, but technology such as 4D imaging requires increased computational performance. Xilinx heterogeneous Zynq UltraScale+ MPSoCs and high-level tool chains are able support both the RF drive, signal capture and image reconstitution, enabling faster scans and wider adoption of 4D.