Design and development of a phantom for tomosynthesis with potential for automated analysis via the cloud

Abstract This paper describes Development of a Phantom for Tomosynthesis with Potential for Automated Analysis via the Cloud. Several studies are underway to investigate the effectiveness of Tomosynthesis Mammographic Image Screening, including the large TMIST project as funded by the National Cancer Institute https://www.cancer.gov/about-cancer/treatment/clinical-trials/nci-supported/tmist. The development of the phantom described in this paper follows initiatives from the FDA, the AAPM TG245 task group, and European Reference Organization (EUREF) for Quality Assured Breast Screening and Diagnostic Services Committee report noting, that no formal endorsement nor recommendation for use has been sought, or granted by any of these groups. This paper reports on the possibility of using this newly developed Tomosynthesis Phantom for Quality Assurance, field testing of image performance, including remote monitoring of DBT system performance, e.g., via transmission over the cloud. The phantom includes tests for: phantom positioning and alignment (important for remote analysis), scan geometry (x and y), chest wall offset, scan slice width and Slice Sensitivity Profile (SSP(z)) slice geometry (slice width), scan slice incrementation (z), z axis geometry bead, low contrast detectability using low contrast spheres, spatial resolution via Point Spread Function (PSF), Image uniformity, Signal to Noise Ratio (SNR), and Contrast to Noise Ratio (CNR) via readings over an Aluminum square. The phantom is designed for use with automated analysis via transmission of images over the cloud and the analysis package includes test of positioning accuracy (roll, pitch, and yaw). Data are shown from several commercial Tomosynthesis Scanners including Fuji, GE, Hologic, IMS‐Giotti, and Siemens; however, the focus of this paper is on phantom design, and not in general aimed at direct commercial comparisons, and wherever possible the identity of the data is anonymized. Results of automated analysis of the phantom are shown, and it is demonstrated that reliable analysis of such a phantom can be achieved remotely, including transmission of data through the cloud.


| MATERIALS AND METHODS
A newly developed Tomosynthesis QA Phantom (Tomophan â , Salem, NY, USA) has been used for testing DBT systems. This Phantom ( Fig. 1) is designed to be responsive to a number of scientific and regulatory groups, including those mentioned in the introduction and other international groups. 6 The phantom is also designed to allow remote analysis via web or cloud. The phantom includes tests for: phantom positioning and alignment (important for remote analysis), scan geometry (x and y), chest wall offset, Slice Sensitivity Profile (SSP (z)), slice geometry (slice width), scan slice incrementation (z), z axis geometry bead, low contrast detectability using low contrast spheres, spatial resolution via Point Spread Function (PSF) and the corresponding Fourier Transform yielding the MTF, Image uniformity, Signal to Noise Ratio (SNR), and Contrast to Noise Ratio (CNR) via readings over the Aluminum square.
The Tomophan (TSP) phantom is comprised of three components: the TSP006 Test Object (see Fig. 2); TSP005, a 14 mm Tissue Spacer; and TSP007, Chest Wall Plate. In the standard configuration, the test components are in the central plane of the phantom. These components can be configured in different positions to allow testing slices in the upper central and lower region of the assembly's 42 mm thickness.
Additionally, the phantom has components so that roll (rotation), pitch, and yaw can be calculated as a method of monitoring or controlling phantom positioning effects (Fig. 3).
The phantom has been used with several commercially available DBT systems (GE, Hologic, Siemens, Fuji, IMS) to study several aspects of image quality performance. The phantom was scanned at the appropriate site and then the image data was uploaded to the Image Owl cloud platform for automated image processing (tomo.imageowl.com).
Results were returned within seconds to the originating site for review by the local medical physics and clinical staff.

2.A | Design of specific test and initial results
In the following section, results are shown for monitoring several of the parameters as listed in Methods and Materials. Data are shown for several commercial Tomosynthesis systems. In future publications, system reproducibility and performance over time, and longitudinal monitoring will also be reported. Examples can also be found on the Image Owl website (tomo.imageowl.com).

3.A | Chest wall offset (missing tissue)
A number of approaches can be used to study the offset, including an independent step wedge. The relevant section of this phantom for determination of chest wall offset (missing tissue) is shown in  nominally contiguous slices. 9 The phantom was used to test typical slice thickness as advocated by each of the five tested vendors A-E, with preliminary data shown in Fig. 8 for slices ranging from about 1 to 5 mm. The data were typically within 0.5 mm or better compared to the nominal slice width from each vendor. On each box, the central mark is the median, the edges of the box are the 25th and 75th percentiles, the whiskers extend to the most extreme data points the algorithm considers to be not outliers, and the outliers are plotted individually (red '+' marker).

3.C | Beads for z axis geometry
A related approach to studying z axis geometry is to study the distances and profiles of beads of known size and known vertical (z axis) spacing. 4 In the current phantom, three 5 mm Aluminum beads are nominally spaced 10 mm apart. The results from superimposing 3 different images, using these 5 mm bead sphere test objects are shown in

3.D | Artifact spread function
The same beads as used in the section Beads for z Axis Geometry can be used for what has been identified as the Artifact Spread Function (ASF) 2,4 is closely related to the SSP in the z axis direction.
The ASF has been determined by calculating the z axis response of a small steel bead. In fact, the ASF was correlated with the convolution of the two-dimensional (2D) point spread function (PSF) of the DBT system and the object function of the bead. 2 In essence, the SSP from examining the sensitivity of adjoining small beads in the bead ramp supplies a function mimicking the results from the ASF. 4 By examining Figs. 9(a)-9(c), it can be seen that as the z axis location of the bead changes, the ASF is generated.
It can be noted that the SSP and ASF can vary not only depending on the x and y [Figs. 9(a) and 9(d)] but also vary depending on the z axis location of the bead as seen in Fig. 7.

3.E | Spatial resolution
The same bead ramps provide a series of point sources (beads) 5,9,10 These beads with their small diameter (18 mm) can be considered  small enough to constitute "points" sources to determine the point spread function (PSF) and resulting MTF 2,5 for current levels of DBT resolution (typically 100 microns or better in DBT mode), particularly when one deconvolves the effective size of the small bead. 11 It can be noted, that strictly speaking, the term MTF should be approached with caution for DBT, because the formal conditions for MTF are not met in systems that may be nonlinear and non-Isoplanatic. 5 Additionally, when one encounters iterative reconstructions in both Computed Tomography (CT), and DBT, the questions of linearity is even further strained and should probably be avoided. This being noted, the term has already been used in several studies 2,10 and will be used in this paper, duly noting these caveats. The "MTF" data shown in this paper is obtained from the Fourier Transform of the PSF. 5 A typical image for the beads (point sources) Fig. 10(a), on the bead ramps (Fig. 6); and the resulting MTF 10 in both the x and y frequency directions is shown in Fig. 10(b). The resulting MTF plots show the known anisotropy between the MTF (x) and MTF (y) results due to the influence of the tube travel direction lowering the MTF in that direction. 12

3.F | Uniformity
A phantom of the same size and background composition as the DBT Phantom is available for checking the uniformity of the DBT response; however, it was decided to investigate the effectiveness of using the multi-purpose phantom for DBT uniformity measurements instead of the dedicated uniformity phantom. The question is whether uniformity can be reasonably measured with the presence of the other test objects. 13 Using the phantom both the regional uniformity and the global uniformity were studied. Uniform ROIs were carefully selected to minimize the effects of other test targets. For this study, we mea- It can be noted that the uniformity measurements as defined in this paper, reveal differences between vendors, as well as trends across the slice dimension. However, these measurements in the multi-purpose Tomophan QA phantom were limited to five 5 mm radius regions. A more detailed uniformity measurement can potentially be carried out using a solid uniformity phantom available as an option to the multipurpose phantom. This optional uniformity phantom is difference than the standard 14 mm tissue spacer that is included in the phantom (as seen in Fig. 2). No significant differences in global uniformity measurements were observed between the multipurpose Tomophan and the uniformity phantom. Therefore, by F I G . 1 1 . Images of Uniformity Section along with designated ROI's for Regional (two large areas), and global uniformity (five smaller ROI's).
F I G . 1 2 . Uniformity variation for top ROI and mid ROI.
| 297 of the noise 6 in a neighboring region; likewise, the CNR can be obtained by taking net signal over an Aluminum square and dividing by the standard deviation of the noise. 6 The Aluminum square is found near the center of the phantom (just inside the circle of low contrast spheres) and is illustrated in Fig. 14 (schematic on left and scan on right).
A typical result (Fig. 15) from one vendor shows the decrease in noise (Standard Deviation) 6 as the mAs is increased and the corresponding increase in CNR. In both cases, the fit equations show the approximate square root dependence of noise on mAs, and the corresponding effect on noise (SD) and CNR. 6

3.H | Low contrast
It can be noted that over and above the standard deviation of the noise, other higher order properties of the noise, such as Noise Power Spectrum (NPS) can be calculated. 6 Thus, the assess-

3.I | Remote analysis via the cloud
As discussed in this paper, all the data can be obtained from reports generated by remote analysis of DBT data from the Tomophan transmitted to analysis software via the cloud. In this paper, the data was provided by Image Owl https://www.imageowl.com/. The analysis can be seen in overview from Fig. 17 and involves the following steps: (a) collection of data, uploading images and data, viewing test results, comparing with QA database, process control precision, system and phantom accuracy of alignment. Examples of report can be obtained from Image Owl.

3.J | Future work
Some initial work has been performed on modeling low contrast performance based on Contrast (C) -Detail (D), C-D curves. 13 For example, one approach involves extracting net (mean signal minus background) signals from circles with diameters equivalent to the diameters of the low-contrast spheres [ Fig. 18(a)]; from the uniform region, next to the low-contrast spheres. Multiple circle means are sampled for each diameter. An example of sampling the values for 6 mm circles (spheres) is shown in Fig. 18(a). This image is from a slice through the phantom centered in a plane with the low contrast F I G . 1 6 . Examples of changes in acquisition parameters on the visualization of these spheres as shown in the four 37mAs images of figure  16; ranging from left of: 60°, CNR of 2.5; 48°, CNR 1.5; 24°, CNR 2.5; and 16°, CNR 1.1. One can also notice the decreased slice thickness with decreased number of views.
Overview of remote analysis via cloud with steps from left to right of: collection of data; uploading images and data; viewing test results; comparing with QA database; process control precision; system and phantom accuracy of alignment. spheres. For each diameter (10, 8, 6, 4, 3, 2, 1.5 and 1 mm), one can then compute the standard deviation (SD) of these means, as shown in Fig. 18(a). In theory, the lower the diameter of the circle (sphere), the higher the SD of the means of the circles (more noise, less precision when using fewer pixels). Reference to other C-D studies, often shows that a hyperbolic model will often fit these points. 11 This seems to hold in the initial results shown in Fig. 18(b) where a hyperbolic fit is made to the resulting noise data as obtained from the noise standard deviation measured over circles of diameter matching the bead diameters.
This area of investigation will be expanded in future work, with more sophisticated models of detection, and possibly ROC analysis. 11 4 | CONCLUSIONS