Topics covered in this article include:
- Measurements Summary
- Phantom Structure
- Analysis Software
- Phantom Scanning
- Analysis Options in QA Settings
- Acquisition Recommendations
- Detailed Discussion of Measurements
The excellent soft-tissue contrast of MRI has motivated strong interest in using MR images to plan radiation therapy. This application of MRI has several significant challenges. Most notable among them is geometric distortion, a critical performance element that must be monitored and controlled. In addition, general quality control protocols are needed that can give assurance that a scanner is performing within specification. These protocols must be efficient, powerful, and capable of being used by practitioners who are not specialists in MRI technology.
The Magphan RT® addresses the general challenge of providing comprehensive Quality Control for an MRI scanner that addresses the needs of the Radiation Therapy Planning (RTP) community. The Magphan RT combines, in a single tool, the ability to measure geometric distortion and a set of measurements that characterize the general performance of an MRI system.
The Magphan RT and associated Image Owl analysis software performs the following measurements:
- Geometric Distortion
- Signal Uniformity
- Laser Alignment
- Signal-to-Noise Ratio
- Slice Thickness (three cardinal axes)
- Resolution (three cardinal axes)
The measurements are completely automated, requiring the user only to drag and drop the DICOM files into the web-based software interface. A detailed report is created, with pass/fail criteria presented first, and detailed information following. The user can set pass/fail criteria for a wide range of tests.
The aim of the Magphan RT is to perform quality control measurements on the same MRI pulse sequences being used clinically, rather than a pre-defined set. Thus the phantom is designed to support a very wide range of image types, with varying contrast and imaging parameters. This approach casts a wider net of quality monitoring than pre-defined protocols because in addition to probing general failures of an MR scanner, it also probes failure mechanisms that may be sequence-dependent, configuration dependent, parameter dependent, and operator dependent.
The Magphan RT performs several measurements with novel techniques that provide detail not normally available in commercial MR quality control tools, including a full 3D signal uniformity measurement, and measurements of the Modulation Transfer Function (MTF), Point-Spread Function (PSF), and Edge Spread Function (ESF).
A more detailed discussion of the phantom measurements is presented in the section 'Detailed Discussion of Measurements.'
The greatest challenge in measuring geometric distortion over large fields of view is making a phantom that covers the field of view but is still of manageable size and weight to be handled easily. A single liquid-filled phantom that covered the fields of view typically desired for RTP would easily weigh more than 100 lbs. The approach taken by the Phantom Lab to overcome this challenge is to make the phantom modular, so that it can be moved in two or three pieces and easily reassembled on the patient table. Proprietary algorithms in the analysis code distinguish between any residual translational/rotational offsets of one module relative to another, and actual geometric distortions in the image. The figure below shows the three-component version of the phantom in its assembled state (top) and partially-assembled state (bottom) to illustrate the modular nature. The maximum weight of any one module is less than 12 kg (26 lbs), making it easily movable by most any scan operator.
There are several advantages to this approach. In addition to the benefits of usability due to the smaller individual components, having a phantom filled with liquid enables many non-distortion measurements to be included without creating complex disturbances to the magnetic field due to the magnetic susceptibility of the phantom materials.
The phantom is filled with a uniform background fill with a T1 relaxation time of approximately 270 ms at 1.5 Tesla. This choice of T1 value is driven by desire to enable the same phantom to be usable for lower field scanners as well as 1.5 Tesla and 3 Tesla scanners. The features within the phantom are all made of plastic polymer materials that appear dark under normal MR pulse sequences. The background fill also contains physiological concentration of saline to provide some degree of loading of the radio-frequency receive coils, and a thickening agent to slow down fluid currents within the phantom.
The features inside consist of the following:
- Fiducial spheres for distortion measurements
- Three separate slice thickness ramps (crossed wedges) to support measurements along any of the three cardinal axes
- Three separate resolution apertures to support detailed resolution measurements along any of the three cardinal axes
- Noise rods for measuring variation in non-signal regions
- Uniform background fill for signal uniformity
The distortion measurements cover a field-of-view of 35 cm x 39 cm x 21 cm (L/R x A/P x H/F) in the three-component configuration, or 35 cm x 27 cm x 21 cm (L/R x A/P x H/F) in the two-component configuration. The phantom can also be rotated 90 degrees to swap the 35 cm dimension and the 21 cm dimension. A work-on-progress will allow the user to upload multiple scans with the phantom translated to provide extended field-of-view measurements.
The Image Owl analysis software has been developed to fit easily into the workflow of a Radiation Therapy clinic. The software is fully automated, requiring nothing of the user beyond dragging and dropping DICOM files into the web-based user interface. The output report is structured to be both immediately useful to the therapist and medical physicist with Pass/Fail summary information presented up front, with links to highly detailed information suitable for analysis by individuals of varying degrees of expertise in MR technology.
The figure below shows the summary section of the analysis software.
Users can set their own criteria for passing each test. The output is categorized into one of three categories: Pass, Warning, or Fail. For measurements that are typically specified within a particular volume, such as distortion and uniformity, separate outputs are created over user-specified volumes.
The report is organized into sections based on the nature of the measurements - Distortion, Resolution, Slice Thickness, Signal-to-Noise Ratio, Signal Uniformity, etc. The figure above shows a few of the outputs from the distortion analysis, including a plot of distortion magnitude vs radius, and a table of distortion measurement statistics.
The phantom is equipped with three sets of alignment markers for positioning the phantom within the scanner. Most of the analysis performed by the service is insensitive to phantom alignment. The Slice Thickness analysis and MTF Analysis, however, require that the phantom be aligned to within two and five degrees of the slice plane, respectively. Below are the requirements to run those tests for each scan plane:
|Slice Thickness Tests
|Yaw < 5°, Pitch < 5°
|Yaw < 2°, Pitch < 2°
|Pitch < 5°, Roll < 5°
|Pitch < 2°, Roll < 2°
|Yaw < 5°, Roll < 5°
|Yaw < 2°, Roll < 2°
The operator places the phantom modules one at a time on the table, sliding one module on top of the prior module using grooved tracks built into the phantom. A locking mechanism helps ensure that the phantom does not come apart easily should it be mishandled inadvertently.
The phantom is designed to support a wide range of MRI pulse sequences. The background fill has a T1 of approximately 270 ms at 1.5 Tesla. The relaxation time is field-strength dependent. This value of T1 was selected to support scanning from 0.3 Tesla up through 3 Tesla. Although a wide range of sequences can be supported, the performance of certain measurements (slice thickness, resolution) are dependent upon adequate signal-to-noise. A configurable threshold will prevent resolution and slice thickness measurements from being performed if the SNR is too low. Certain severe artifacts can also interfere with the analysis of the phantom.
The complete phantom should be scanned, with coverage extending at least a few mm beyond the edge of all fiducial spheres. The orientation of the scan plane can be in any direction, however the resolution and slice thickness measurements will be performed only for the 'cardinal' orientations of the scan plane (axial, sagittal, or coronal). The phantom must be aligned within two degrees for these measurements to be performed. It is very easy to achieve alignment significantly better than two degrees by using the laser alignment marks.
The phantom is designed to best support slice thicknesses in the range of 1-3 mm. Beyond 3 mm, the fiducials may not be located properly within the images. The slices should be contiguous so that there are no gaps between the slices.
Analysis Options in QA Settings
There are several analysis options for Magphan phantom. We recommend contacting support before adjusting any of the settings not mentioned in this article.
Typically the only setting we recommend changing is the Sphere Fraction Detection Threshold for sites with small artifacts. The Sphere Fraction Detection Threshold is a sanity check to ensure a certain number of spheres are found before running the distortion analysis. The default is 0.97 or 97% of the expected spheres. The default value is rather conservative. For scanners with low distortion, the setting may be safely lowered.
To adjust the analysis options, navigate to the "Manage Schedule Configurations" page, hover over the desired schedule, and click the QA Settings button (it appears as a gear) for that schedule. After any needed changes, be sure to click the Save Machine QA Settings button to save your changes.
Note: the settings can be reset to default at any time by clicking the Reset to Default button.
Slice Thickness Analysis Options Section
- SNR Threshold variation with slice thickness - this option adjusts the necessary "Slice Thickness SNR threshold" for the slice thickness analysis to be run, so that different nominal slice thickness values require different "Slice Thickness SNR Threshold" in order for the slice thickness analysis to run. The need for this option arises because thicker slices require a higher Signal to Noise Ratio (SNR) in order for the slice thickness ramp feature to perform a quality measurement. Quantitatively, the required SNR increases by the value of SNR Threshold variation with slice thickness for every millimeter increase in slice thickness, relative to a baseline slice thickness of 1.5 mm. For example, if SNR Threshold variation with slice thickness is set to 5, then for every millimeter increase in slice thickness relative to 1.5 mm, the required SNR increases by 5. The required SNR is given by the formula SNR = Slice Thickness SNR threshold + SNR Threshold variation with slice thickness * (nominal slice thickness - 1.5mm), where 'nominal slice thickness' refers to the slice thickness selected by the user for the acquisition.
Segmenter Options Section
Remove dim fiducial spheres - this option, when set to 1, causes the analysis to exclude fiducial spheres in regions where the image intensity is much more dim than it is for most of the phantom. The typical use case of the option is to prevent ghosted copies of spheres from being detected. To disable this feature, set it to 0. The threshold for cutoff is determined by the analysis option brightness_cutoff_ratio.
Cutoff ratio for dim signal near fiducial spheres - this option is active when "Remove dim fiducial spheres" is set to 1. It controls the threshold for image intensity in a region near a sphere that causes the sphere to be excluded. Quantitatively, it is the ratio of the intensity of the image around the sphere to the average intensity around the other spheres in the phantom that have been found. For example, if " the Cutoff ratio for dim signal near fiducial spheres" option is set to 10, then spheres that are in regions with intensity 10 times less than the average will not be detected by the analysis. The purpose of the feature is to prevent dim, ghosted copies of spheres from leading to false sphere detections.
The design aim for the Magphan is to support actual clinical sequences as is feasible, rather than prescribing a fixed quality control sequence. While the phantom may not perform well on all sequences, it works well over most ranges. It is recommend to start by trying the clinical sequences in use, particularly those that are not T2 weighted, or some specialized sequence like diffusion imaging, and see if it processes.
Some T2 sequences work fine. There are at times banding artifacts that may come in and render some the spheres undetected, but those images may still be processed. It may require adjusting the QA setting that allows the analysis to process with fewer spheres. By default, 97% of the spheres must be detected for the distortion analysis to run, but this threshold may be changed in QA settings. The caveat is that if too many are missing, the distortion measurement could become unreliable. In our experience, on low-distortion scanners such as what you most likely have, the distortion analysis performs well even if up to 8 or 9% of the spheres are missing, and if you lower the threshold to allow it to run through that, the measurements are still useful. If you are experimenting with T2 weighted sequences and want to work with this, adjust the Sphere Fraction Detection Threshold in the QA settings.
Detailed Discussion of Measurements
At the core of any MR distortion measurement is a fiducial marker that gives a precise location of a single point. Point-like sources are not practical in MRI because the signal is proportional to the volume of the source. Furthermore, in the presence of noise, the optimal way to localize an object is to use a larger object rather than a smaller one, as the location of the larger object is less influenced by noise in the scan due to the increased averaging throughout the volume (or around the edges) of the feature. The Magphan RT uses spheres of diameter 1 cm as the primary fiducial marker for the distortion measurement. One can think of the measurement as an edge detection of the boundary of the sphere, with the location returned by the measurement the center location of the sphere.
Another important design choice is the spatial density of the fiducial measurements. MR distortions are by nature slowly varying, and a high sampling density is not necessary to accurately characterize the distortion field. This can be thought of in a similar manner to discrete sampling of a bandwidth-limited continuous signal. In the case of periodic sampling, one needs only to sample a continuous signal at the Nyquist frequency or higher to get complete information about the continuous signal. A slowly-varying signal can be accurately characterized by interpolating appropriately between discrete measurements.
By setting the sampling density appropriately, the phantom can serve multiple purposes other than distortion measurement. The Magphan RT uses a sampling spacing of approximately four cm for the fiducial markers, leaving room in between fiducials to make a more complete measurement tool, capable of measuring other image quality metrics such as resolution, slice thickness, and signal uniformity, off a single MR acquisition.
One of the key technologies in the Magphan RT is the multi-module design that enables the phantom to be handled easily. The challenge imposed by such a design is to properly distinguish between any residual translational or rotational offsets ('rigid-body' transformations) and an actual geometric distortion of the image. The Image Owl software is able to perform this separation by relying on the fact that a uniform rigid-body transformation of one module relative to another is not a possible distortion pattern that can be imposed by an MR scanner. The distortions imposed by an MR scanner ultimately come from disturbances in the magnetic field pattern, which must satisfy Maxwell's Equations.
Thus the Magphan RT software is able to determine the underlying distortion field by allowing, for each module, separate rigid-body degrees of freedom in the displacement field, and assigning the distortion to be the non-rigid-body portion of the displacement measurements. This algorithm lies at the core of the multi-component technology of the Magphan RT, and eliminates any residual rigid-body offsets of one module relative to another. The technique can also be used to combine acquisitions from multiple MRI scans to perform an extended-field-of-view measurement, which is a work in progress for the product.
Distortion Measurement Outputs
The distortion measurements currently output a plot of the distortion magnitude vs radius, statistics on maximum distortion over two user-set diameters, and a 3x3 detailed distortion plot that can be used to ascertain information about the 3D distortion pattern of any component of the distortion. The diameters over which the distortion measurement are performed can be selected in the QA Settings menu. Each of these outputs is discussed below:
Distortion Magnitude vs. Radius
One of the simpler graphical outputs that is useful for characterizing distortion is a plot of the magnitude of the distortion vector vs the location at which that distortion was measured. In the plot below, the vertical axis depicts the magnitude in mm and the horizontal axis depicts the distance from isocenter (radius) where the measurement occurred, also in mm. The plot can be useful for getting a quick understanding of where the distortion stays within various limits.
Fiducial Location Statistics Tables
Although the graphical output contains a large amount of information, for tracking and automated failure detection purposes it can be more useful to set a specification on a specific measurement. The Fiducial Location Statistics table presents measurements within the user-specified diameter.
The measurements presented are:
- Max Distortion Magnitude - the magnitude of the largest distortion measured within a given diameter
- Mean Distortion Mag Top 10% - the mean of the magnitude of the distortion of the 10% most distorted measurements. This measurement presents a somewhat more broad view of the high-end distortion than the single largest distortion measured.
- Max X,Y, and Z distortions - the maximum distortion in each of three directions. These are the vector components of the distortion in the Left-Right, Anterior-Posterior, and Head-Foot directions, respectively.
- Mean X,Y, and Z distortions Top 10% - the mean of the component of distortion in each direction, with the average taken over the 10% largest measurements (in each respective direction)
Fit Distortions Plot
The Fit Distortions Plot (also referred to as the Distortion 3x3 plot) depicts the full vector information of the distortion measurements over the 3D volume. Because distortion is a vector, and the location where the distortion is measured is also a vector, there are nine separate plots that describe this data.
In the figure above, each column represents one component of distortion, and each row represents the location of the distortion measurement in patient (Left, Posterior, Superior) coordinates. The horizontal axis shows the magnitude of the distortion, and the vertical axis shows the location of the measurement along a particular direction. For example, the upper left plot shows the distortion in the Left-Right direction on the horizontal axis and the location of the measurement in the Left-Right direction on the vertical axis. In an idealized distortion-free scanner, the points would all line up on a vertical line at 0 distortion. Each circle represents the measurement of a single fiducial sphere, with the light blue color highlighting the 10% largest distortions.
Although this figure is complex, it reflects the inherent complexity in the underlying data it represents. The distortion is a three-component vector function over a three dimensional space, meaning that nine separate plots are required to depict the full information. This representation contains a great deal of information about the distortion in a single figure. As another example, the second plot from the top in the third column represents the distortion in the Head-Foot direction as a function of the Anterior-Posterior location of the measurement. In this coordinate system, the positive directions follow the convention of the DICOM standard, which is Left-Posterior-Superior.
Fit Distortion Measurement Download
Along with the above outputs, the user can download a .CSV file containing the full set of distortion measurements, including the locations and vector distortions at each location. This file is named 'FitDistortionPoints.csv'.
The FitDistortionPoints.csv describes the distortion measured at each fiducial sphere location in the phantom.
Column 1: an integer that identifies the id of the module from which the sphere came
Column 2: an integer that identifies the id of a sphere within its module
Columns 3-5: The locations in Left, Posterior, Superior (LPS) coordinates of the detected location of the center of each sphere within the DICOM image.
Columns 6-8: The LPS coordinates of the fit distortion vector at that sphere location.
The fit distortion vector is defined as the displacement of the detected location from the true location of a fiducial. The true location, then would be determined by subtracting the fit distortion vector from the detected location.
Resolution of an MR system is measured in the Magphan RT using a circular aperture of 2-cm diameter. There are three orthogonal aperture features, so that if the acquisition plane falls along any of the three cardinal axes (axial, sagittal, coronal), the correct feature is used for the calculation. The phantom must be aligned within two degrees of the scan plane for the analysis to run. Otherwise, a warning is issued and the analysis is not performed.
The interior of the aperture is filled with fluid and hence appears bright in the MR image, and the exterior appears dark. In an ideal imaging system with infinite resolution, the aperture would appear as a feature with a sudden signal change at the perimeter of the aperture. Due to the scanner's finite resolution, the edges of the aperture get blurred. The discrete sampling then produces a blurred, pixelated image, as shown in the figure below.
The complete mathematical description of resolution is contained in the system Point Spread Function (PSF). It is often more convenient, however to observe derived quantities such as the Modulation Transfer Function (MTF), or Edge Spread Function (ESF). The analysis calculates all of these quantities and presents certain characteristic numbers that can be useful for quantitative comparison and tracking. The analysis is capable of calculating these quantities separately for the row and column directions. For scans with sufficiently high signal-to-noise, this can be a reliable measurement. For scans that have lower signal-to-noise ratios, and where isotropic resolution is expected, it can be advantageous to calculate isotropic MTF, PSF, and ESF functions. This setting is available in the QA settings menu.
Modulation Transfer Function (MTF)
The Modulation Transfer Function is the magnitude of the Fourier Transform of the Point Spread Function. It can be useful to observe the MTF to understand the nature of the filtering that is being applied to the image.
Point Spread Function (PSF)
The Point Spread Function can be viewed as the fundamental and complete description of the resolution of an imaging system. From this function, all other resolution measurements can be derived, including low-contrast resolution. As such, it is an output that many may be interested in observing, with a simple scalar characterization being the Full Width at Half Maximum Value (FWHM). The plot of the Points Spread Function and the FWHM value in each direction is depicted in this measurement output. In the plot below, the circular marks represent the locations of nearby pixels.
Edge Spread Function (ESF)
The Edge Spread Function is arguably the most useful characterization of resolution for an MR system, since the boundary of most features being observed resembles an edge more than it resembles a point. The ESF measures the spreading that would occur if an ideal step transition edge were imaged. Instead of a sharp edge in the image, the edge appears to be spread out, and some ringing is usually introduced on either side of the transition.
A useful scalar characterization of the resolution is the width in mm required to transition from 10% to 90% of the step magnitude. This number is called the Edge Spread Function 10-90 Transition Width, and is shown in the table underneath the plots of the Edge Spread Function:
The signal-to-noise ratio (SNR) is calculated by taking the average of the signal at a region near the center portion of the phantom, and dividing it by the standard deviation of the signal in a dark region of the image. The dark region is taken as the composite region consisting of the Noise Rod features of the phantom, which are cylinders of solid plastic from which there is no MR signal generated.
Slice Thickness Measurement
The slice thickness is measured in a conventional way using crossed ramps. The ramp region is a 1-mm thick slot cut out of solid plastic, so that the slot appears bright in an MR image. The slice thickness is calculated from the full-width-half maximum value of the signal region. The crossed ramps compensate to first order for angular misalignment, and the final slice thickness is calculated from the geometry of the ramps. An ROI of the ramp region is shown below:
It is normal for most pulse sequences that this slice thickness comes out substantially larger than the nominal slice thickness, particularly for sequences that do not employ multiple refocusing RFpulses, and for thinner slices.
There are three orthogonal slice thickness features, so that if the MR acquisition is performed in any of the three cardinal orientations (axial, coronal, sagittal), the pertinent ramp is identified and used for the calculation. The slice thickness ramp must be aligned with two degrees of the scan plane, or else the analysis will not be performed and a warning will be issued by the analysis service.
Signal Uniformity is an important measure of image quality, as changes in the uniformity can reflect a problem with a phased array coil. The analysis performs uniformity measurements over three dimensions by using the background regions that consist only of background fluid, and no features. The user can select two diameters over which to measure the uniformity.
Two characterizations of uniformity are provided:
- The 'normalized standard deviation,' which is the standard deviation of the signal within the specified diameter divided by the mean of the signal in that diameter
- The 'normalized spread,' which is the difference in signal intensity between the mean of the 10% brightest voxels and the 10% dimmest voxels within the specified diameter, divided by the mean of the signal within that diameter.
The mean signal is also reported, though this is not itself a very meaningful image quality metric. However, if it were to change drastically for what is nominally the same acquisition, it could reflect a problem with the scanner.
Also shown in the output are histograms of the signal intensity within each spherical diameter.
Below are examples of the histogram plot and the uniformity statistics table. The diameters over which the analysis are run can be selected in the QA settings menu for the schedule.
Laser Alignment (Phantom Placement)
The relative alignment between the MR system isocenter and the laser alignment system is monitored with the Phantom Placement Test. The analysis determines the translational and rotational offsets of the phantom relative to the MR system isocenter, and reports them as tables.
The Translational Offsets table describes the translation vector that points from the system isocenter to the center of the phantom, reporting the components in the patient Left-Right, Anterior-Posterior, and Head-Foot directions. The Angular Deviation table describes the rotational offset of the phantom relative to isocenter using Yaw, Pitch, and Roll. Examples of these tables are shown below.