This educational talk will discuss image reconstruction software platforms that go beyond the platforms provided by the vendors. These platforms can be used for prototyping new reconstruction methods and for enabling capabilities such as real-time interactive scanning. Open-source image reconstruction software libraries accelerate image reconstruction research and enable reproducible research.
Highlights
Target Audience
Research scientists, engineers, and clinicians interested in a discussion of alternative image reconstruction platforms and how they accelerate image reconstruction research, provide new capabilities, and enable reproducible research.
Outcome/Objectives
To understand the potential advantages of alternative reconstruction workflows. including rapid prototyping, providing real-time interactive capabilities, building on software libraries to accelerate image reconstruction research, and enabling reproducible research.Overview
MRI scanner vendors have well-developed image reconstruction software platforms that are essential for clinical MRI. These platforms are tightly integrated with data acquisition, image display, image analysis, and image archiving. So, why should anyone consider an alternative software platform, when the vendors platforms are so capable? There is a long history of image reconstruction research being conducted using alternative platforms, and there are a number of potential advantages of using them. Two resources for exploring these platforms are ISMRM’s MR-Hub [1] and Open Source Imaging [2]. This talk will explore motivations for using alternative platforms, show some examples, and discuss how these platforms can accelerate image reconstruction research and enable reproducible research.
Prototyping
One practical reason for using an alternative reconstruction platform is that academic researchers do not always have full access to programming the image reconstruction platforms of the vendors, although that is less of an issue now than in the past. Still, it is often useful to prototype a new image reconstruction method using a familiar general purpose software platform, such as MATLAB or Python. Many image reconstruction researchers do not have a research agreement with a vendor, and they make significant contributions to the field using off-line reconstruction. Even with full access to a vendor’s platform, prototyping a method off line is often useful. Once the method is fully developed and shown to be useful, it can be ported to the on-line platform for clinical studies or commercial availability.Expanded Capabilities
The vendor platforms are designed to support their current products and projected future capabilities. Researchers are often interested in pushing the envelope, and thus they may need to write their own platform or use one provided by someone else. One early example of this was spiral k-space scanning [3]. Understandably, the vendors did not initially provide specific support for non-Cartesian image reconstruction. The solution we used in [3] was to have the scanner save the raw data to disk and then to reconstruct the images using a gridding image reconstruction program written in C. We wrote scripts to automatically call the reconstruction program, add a DICOM header to the images, and import the images into the scanner’s database. The reconstruction program typically ran on a separate network workstation, although it could run locally on the scanner’s host computer. Many other researchers have used similar solutions for various types of image reconstruction.
Another example of additional capabilities is real-time image reconstruction, which enables a variety of applications such as image-guided therapy and real-time cardiac imaging. This typically involves transferring the acquired data from an embedded computer on the scanner to a workstation after each readout. An early implementation used a custom interface board and a dedicated array processor [4], and was then extended to real-time interactive system [5]. Another real-time interactive implementation used a real-time workstation to access the data in real time using existing scanner hardware with custom software, and then the workstation distributed the data to a networked cluster of workstations to reconstruct temporal frames in parallel [6]. This system was capable of both Cartesian and non-Cartesian real-time image reconstruction. It was used for various cardiac patient studies [7], and during these studies, every workstation in the lab was commandeered for real-time computation. Of course, greater computational capabilities are now available upon a single workstation, which have enabled various implementations of more computationally-intensive accelerated reconstructions, using either a single workstation [8,9] or a computing cluster [10].
Software Libraries and Reproducible Research
Open-source software libraries substantially accelerate image reconstruction research, enabling researchers to build upon the efforts of others. A partial list of available libraries is given at MRI Unbound [11], which is a predecessor of MR-Hub [1]. Widely used libraries include those with fundamental reconstruction tools [12], compressed sensing methods [13], development environments [14], and real-time platforms that interface directly to the scanner [15]. As MR image reconstruction research incorporates more machine learning methods, it is building upon the extensive software libraries available in the machine learning community, such as TensorFlow [16]. Anyone with a software library who would like to make it more widely known is encouraged to submit it to MR-Hub. Submission instructions are on the website.
There is currently a movement toward making research more reproducible, and this is increasingly encouraged by funding agencies, such as the NIH [17]. For image reconstruction research, this often means making all of the code and data associated with a publication available. The ISMRM Raw Data Format (ISMRD) [18] is an important tool for enabling cross-platform sharing of code and data. In 2017, the ISMRM Reproducible Research Study Group [19] was formed to promote reproducible research, and new members are welcome.