CA2534701A1 - High energy soft tissue imaging using diffraction enhanced imaging - Google Patents

Description

HIGH ENERGY SOFT TISSUE IMAGING USING DIFFRACTION
ENHANCED IMAGING

FIELD OF THE INVENTION
The present invention relates to the field of medical imaging, and more particularly to a new method and apparatus for multiple-image radiography.
According to a preferred embodiment, a method of high energy soft tissue imaging using diffraction enhanced imaging is provided.

SUMMARY OF THE INVENTION
A design and feasibility study for a new imaging method is disclosed, called multiple-image radiography (MIR), which we hope will become a viable alternative to conventional mammography. Like today's mammography, MIR uses x-rays to produce an absorption image, but MIR also produces two new kinds of images, one showing refraction effects, and one depicting ultra-small-angle scattering (USAXS).
The USAXS
image, which is completely unique to MIR, indicates the presence of textural structure of tissue on the scale of microns. It appears that USAXS may be a characteristic and clinically important signature of spiculations and calcifications in the breast, a hypothesis that will be examined in this project. The refraction image provides a highly detailed representation of soft-tissue structures, and appears to be especially effective for visualizing masses and spiculations.

MIR offers numerous potential advantages over conventional mammography: 1) the refraction and USAXS images provide excellent contrast for visualization of soft tissue, especially masses, fibrils, and calcifications; 2) all three MIR
images are virtually immune to the scatter degradation that is characteristic of conventional radiography; 3) MIR may lessen the need for breast compression during imaging; and 4) MIR can dramatically reduce patient dose because it can work at high x-ray energies where absorption is low.

The key challenge for MIR is that it currently can be performed only using synchrotron sources of x-rays. To make MIR clinically usefiul, we must eventually build a compact M IR i mager u sing c onventional x -ray t ubes o r o ther compact so urces. 0 ne preliminary study suggests that MIR can yield dramatically lower patient dose than a - I -conventional mammogram (by a factor of 250-750) due to its ability to function at higher x-ray energy. The images thus produced would have roughly 1/3 to 1/9 the x-ray fluence at the detector as a conventional mammogram, but would still have around 3-19 times the contrast-to-noise ratio of conventional mammography. Therefore, improved image quality and greatly reduced patient dose are potential benefits of MIR.
There are additional causes for optimism about this result: 1) the system simulated in the software can be further optimized; and 2) several avenues to improve x-ray throughput, which are outlined in the proposal, are not included in the simulated scenario.

We anticipate that tungsten Kal (59.3keV) will be the preferred energy, because of lower absorption (higher detector fluence), and because commercial tungsten tubes are readily available. However, the signal-to-noise ratio of MIR is better at lower energies.
Therefore, to be complete, we feel we should study various energies on the continuum, and we chose to sample uniformly from 30keV-60keV in l OkeV increments. If one of the lower energies is found to produce greatly superior results, this may drive future investigations of unusual sources. For example, barium hexaboride targets produce x-rays at about 30keV, samarium's Kal line is at about 40keV, and tabletop synchrotrons under development at Lyncean Technologies, Inc., can produce any energy in the range of interest. Tubes with exotic targets, and tabletop synchrotrons, may look impractical or prohibitively costly today, but we feel it is appropriate to investigate these options in the proposed feasibility study, along with practical options that could be implemented with current technologies. The energy dependence of detector response and spatial resolution will need to be taken into account in the proposed computer simulations, and we fully intend to do so. In the synchrotron-based studies, energy dependencies will be inherent in the experimental results that we will obtain. We have performed MIR more properly with a digital detector, and the new results are shown in Figures 1 and 7. The USAXS
images now show information with very high sensitivity to fibrils and calcification.

~ ~~

DETAILED DESCRIPTION OF THE DRAWINGS
Reference is made herein to the appended drawings. Further features and advantages of the present invention will become apparent from the following detailed description, taken in combination with the appended drawings.

Figure 1. Conventional digital mammogram (left) of infiltrating ductal carcinoma, showing region of interest. Images below show magnified view of region of interest:
conventional m ammogram (below 1 eft) and M IR i mages (below). T he s piculations are highly visible in the MIR refraction and USAXS images, but much less so in the MIR
absorption image, and almost invisible in the conventional mammogram. This suggests that refraction/USAXS contrast may be much more important than absorption contrast for visualizing spiculations. The color composite image (below right) shows the USAXS
image as a color overlay on the refraction image, emphasizing that the USAXS
image appears to highlight spiculations. The detailed visualization of spiculations provided by MIR may be an important tool for surgeons to determine extent of tumor proliferation.
Figure 2. MIR images with radiograph for comparison.
Figure 3. MIR imaging system (not drawn to scale).

Figure 4. Effect of object on a perfectly collimated beam, illustrating the physical effects measured by MIR..

Figure 5. Examples of the deconvolved object function f(6; x, y) as a function of angle, at representative points in the phantom.

Figure 6. Radiograph of breast tissue with invasive lobular carcinoma, showing region of interest. DEI images and peak image show much greater contrast for fibrils than conventional radiograph at upper right.
Figure 7. Radiograph and MIR images of large benign calcifications in an uncompressed whole breast specimen. MIR-USAXS looks similar to MIR-absorption, but can be obtained at high x-ray energy, because it is based on refraction contrast. The bottom right image is a composite of the refraction and USAXS images.
Figure 8. Microcalcifications in uncompressed specimen.

Figure 9. Optical layout for simulated MIR system: (a) schematic showing main elements, (b) and (c) are scale drawings of top and side views, respectively.
Note that the distance from source to detector is set to 1 meter.

~ ~ ' Figure 10. Schematic diagram of detector envisioned in simulation study.
Direct x-ray to charge detectors allow good spatial resolution and stopping power at high x-ray energy, such as that produced by a tungsten CT tube.
Figure 11. Early laboratory-sized DEI imager (left) and test images of a Lucite rod partially immersed in water (right).
Figure 12. MIR imaging system at synchrotron at Brookhaven National Laboratory.
Figure 13. Simple detectability model: imaging a blood vessel in tissue.
Figure 14. M IR images (top row) and DEI images (bottom row) based on simulated photon noise with maximum count in entire data set of only 50 counts/pixel .
Figure 15. Noise standard deviation vs. bias for various noise smoothing methods. Good noise reduction can be achieved in MIR while preserving low bias.
Figure 16. Plot of USAXS parameter vs. object thickness, demonstrating linearity.
Figure 17. MICT phantom.
Figure 18. MICT images of phantom in Fig. 17.
Figure 19. MICT images of bone specimen.
Figure 20. Conceptual block diagram of system concept to be optimized in the proposed research.

Figure 21. Recovery of refractive-index distribution from noisy data by direct inversion (lower left) and LMMSE estimation (lower right).
Figure 2 2. P erspective view of a logspiral focusing element coupled to the first DEI
crystal (monochromator).
Figure 23. Illustration of focusing effect of log-spiral element, with source at caustic.
Figure 24. Characterization system for experimental studies, which will be constructed on an optical table in the laboratory.
Figure 25. Scanning motions for MICT.
- 1+

DETAILED DESCRIPTION OF THE INVENTION
A. SPECIFIC AIMS
In 1997, members of our research team proposed a new type of phase-sensitive x-ray imaging, called diffraction-enhanced imaging (DEI) [1]. In 2002, we proposed a second generation of this technique, called multiple-image radiography (MIR) [2]-[5], which represents a significant advance for this technology. Both methods have generated a great deal of interest, and are now being studied by research groups worldwide (e.g., [6]-[11]).
In concept, MIR represents a dramatic improvement on conventional radiography.
Whereas a conventional radiograph measures only x-ray absorption, MIR simultaneously produces three parametric images, which separately depict: 1) x-ray absorption, 2) x-ray refraction, and 3) ultra-small-angle x-ray scattering (USAXS). In addition, all three images are virtually immune to the scatter that degrades conventional radiographs; therefore, MIR images show fine image detail not seen in today's clinical mammography.
MIR produces impressive images of soft tissue at low dose, and thus offers tremendous promise for improving specificity and reducing interpretation errors in clinical mammography. Preliminary evidence suggests that refraction- and USAXS-based images exhibit dramatically greater contrast than radiographs for breast cancer features, most notably small fibrils (spiculations) that indicate that a tumor has extended into surrounding normal tissue [12].
MIR images are virtually immune from scatter, and can be obtained at high x-ray energy where absorption is low. Therefore, MIR may lessen the need for significant breast compression during imaging, a hypothesis that is currently being tested under a separate project [13]. MIR may also lead to significant reduction of patient dose, due to its ability to perform well at high x-ray energy.
A clinical MIR or DEI device for breast imaging has not yet been designed or investigated. Thus far, DEI
and MIR studies have been conducted entirely using synchrotron light sources, which are powerful research tools, but have no relevance for routine clinical use. Without question, the next step for this technology is to tackle the technical challenges involved in transferring this synchrotron-based technique to the clinic, and to demonstrate its diagnostic importance. Our research team, which includes experienced engineers, physicists, and clinicians, including the original inventors of both DEI and MIR, is well placed to resolve the overarching issues of technical feasibility and clinical utility.
The principal technical challenge for MIR is that, in the process of preparing the x-ray beam, the beam is greatly attenuated; therefore, the initial x-ray source must be very bright.
Therefore, if a clinical device is constructed by simply replacing the synchrotron with an x-ray tube, the resulting images would be essentially the same, but potentially very noisy. In this project we will seek to overcome this problem by optimizing the hardware design, image-processing methodology, and data-acquisition parameters of MIR.
Therefore, in this project we will: a) seek to determine how breast lesions are visualized in MIR images; b) predict the best image quality that might realistically be obtained from a compact MIR mammography system; and c) use radiologist reader studies of pathology-proven mastectomy specimens to compare realistic MIR images (emulated using a synchrotron) with conventional mammograms. A byproduct of this research will be candidate designs for a compact MIR
breast imager, which will lay the groundwork for a future project to construct the first-ever clinical MIR
prototype.
The specific aims of the proposed project are as follows.
1. To determine the characteristic features of breast cancer in MIR images, explain and quantify the physical contrast mechanisms giving rise to these features, and develop a set of realistic MIR breast phantoms.

2. To predict the maximum performance that can be achieved by a clinical MIR
imaging system through optimization of the imaging hardware components and design, the image-processing algorithms upon which MIR
is based, and the data-acquisition parameters. In this effort, we will use the synchrotron at Broolchaven National Laboratory as a physically realistic simulator of the images that might be produced by a future compact MIR
device.

3. To perform expert radiologist reader studies to determine whether images obtainable from a clinical MIR system will be significantly better than conventional digital radiographs in terms of visualization of breast cancer features.

4. As a byproduct of Aim 2, compile complete specifications of several optimized designs for a clinical MIR breast imager, in preparation for a future follow-on project to construct the first prototype system of its kind.

__15-B. BACKGROUND AND SIGNIFICANCE
B.1 What is MIR?
Multiple-image radiography (MIR) is a phase-sensitive x-ray imaging modality, originally developed by our research team [2]-[4]. Unlike a conventional radiograph, which produces a single image showing x-ray absorption, MIR
simultaneously produces three different images: 1) x-ray absorption, 2) x-ray refraction, and 3) ultra-small-angle x-ray scattering (USAXS). The MIR absorption image shows the same information as a conventional radiograph, except that it is virtually free of scatter degradation. The refraction and USAXS images show, respectively, the effects of large-scale (>50 m) and small-scale (-1-50 m) variations in the refractive index of the object. Thus, MIR
is well suited to soft-tissue imaging, and very promising for mammography.
The goal of this research is to investigate the feasibility of bringing MIR
into routine clinical application.
Specifically, we will aim to determine, under practical technical constraints, whether a clinical MIR imager can outperform conventional radiography in terms of diagnostic information. To explain our motivation for this investigation, let us begin by explaining the potential importance of MIR for breast cancer imaging.
B.2 Significance for women's health B.2.1 Expected benefits of MIR
Development of a clinical MIR imager may have major significance for women's health for the following reasons:
1) as we will show later in the proposal, MIR is expected to produce very high contrast for the features that are most ~ ~ ~

important to detection and characterization of breast cancer; 2) by allowing imaging at high x-ray energies, MIR may permit reductions in absorbed patient dose; and 3) because scatter degradation is essentially eliminated by MIR, and because of reduced absorption and dose, MIR may reduce the need for breast compression during imaging, which is a painful part of today's clinical procedure.
Our preliminary results, such as those shown in Figure 1, suggest that very fine image details, such as fibrils (also known as spiculations) emanating from cancerous masses, can be seen in greater detail in MIR images than in conventional radiographs. This suggests the following potential clinical benefits of MIR: 1) MIR may permit cancer to be detected at an earlier stage than is possible with conventional radiography, 2) MIR may improve specificity (i.e., correct discrimination between malignant and benign lesions), and 3) clear visualization of the extent of fibril proliferation may aid in surgical planning.
B.2.2 The need for improvements in clinical breast imaging Screen-film mammography has been studied extensively for the last 40 years, and because of many large randomized screening trials, it is known to reduce breast cancer mortality by approximately 18-30% [14]. The rate of breast cancer death in the last few years has begun to decline, likely due in part to the widespread use of this imaging test [15]. However, standard screen-film mammography is neither perfectly sensitive nor highly specific. Dense breast tissue and diffuse involvement of the breast with tumor tends to reduce the sensitivity of screening mammography [16]. Approximately 10-20% of breast cancers that are detected by self-examination or physical examination are not visible by screen-film mammography [17].
In addition, when lesions are detected by mammography and biopsy is recommended by experienced radiologists, only 5-40% of lesions prove to be malignant [18].
Furthermore, approximately 30% of breast cancers are visible in retrospect on prior mammograms [19]. Digital mammography is under investigation, but early studies do not show this technology to represent a significant improvement over film mammography [20].
Other technologies are also under study, namely breast magnetic-resonance imaging (MRI) and sonography, but their large-scale and long-term effects on breast cancer mortality have not yet been measured through multi-center randomized clinical trials [21],[22].
B.3 Technical introduction to MIR
MIR is an x-ray imaging method that simultaneously produces three images of an object:
absorption, refraction, and USAXS. Figure 2 shows a simple phantom that illustrates these three effects. The phantom consists of a Lucite rod (Icm diameter) in front of eight sheets of paper arranged to form a step wedge, varying in thickness from one sheet (at the right end) to eight sheets (at the left end). In Figure 2, the MIR absorption (attenuation) image is similar to a conventional radiograph, but exhibits much greater contrast owing to scatter rejection. The MIR refraction-angle image depicts the magnitude of small beam deflections caused by large-scale refractive-index features. Thus, only the Lucite rod shows up strongly in the refraction image. The MIR USAXS iinage quantifies angular divergence of the beam caused by the presence of textural structure within the object at a scale smaller than a pixel (<50 m). In this phantom, the USAXS image shows the fibrous structure of the paper.
MIR is based on images acquired using a specialized x-ray optical setup called a Bonse-Hart camera [23], shown in Figure 3. This device uses three perfect diffracting crystals as a highly selective angular filter, which allows the angular content of an x-ray beam to be precisely analyzed, with angular precision on the order of 1 -- ~ ...-microradian. By rotating the third crystal, called the analyzer, one can measure the amount of radiation traveling in a particular direction. In reality, of course, the system responds to a small range of angles. The angular sensitivity function of the crystal system is called the intrinsic rocking curve R(B), which plays the same role in the angular domain as a point-spread function plays in the spatial domain.
Thus, as we have shown in [4], images acquired as a function of analyzer position 0 can be modeled by the following convolution: g(B; x, y) = R(B) * f(B; x, y) , where f(B; x, y) is a function that describes the effect that the object would have on a perfectly collimated x-ray beam. The function f(B; x, y) captures several important properties of the object, especially absorption, refraction, and USAXS.
To understand why this is so, consider Figure 4, which shows an idealized experiment in which the object is illuminated with a perfectly collimated beam. Here, we focus our attention on only a small portion of the beam destined for location (x,y) in the image. This portion of the beam can be deflected by refraction, broadened into a cone by USAXS, and reduced in amplitude by absorption. (In the figure, the angles are dramatically exaggerated for illustration purposes. On a macroscopic scale, the transmitted beam can be regarded as remaining essentially collimated).
Characteristic examples of 9-profiles of f(B; x, y) are shown in Figure 5 for various points in the rod-and-paper phantom, which illustrates how various combinations of absorption, refraction, and USAXS can attenuate, shift, and broaden the function f(9; x, y) . Note that, when no object is present, f(B; x, y) is ideally a Dirac delta function.
One can describe the object's projected image properties at (x, y) in terms of the following three parameters. 1) Absorption is determined by the loss in beam intensity, from which the absorption coefficient can be computed by inverting Beer's law. Absorption is determined by the zeroth moment (integral) of f(B; x, y) over B. 2) Refraction is defined as the mean angular deflection of the beam at each pixel (i.e., the first moment (centroid) of f(B; x, y) ). 3) USAXS is defined as the width of the beam's cone angle at each pixel, as measured by the second central moment of f(B; x, y) , which is equivalent to statistical variance of the photons' deflection angles. These are the quantities depicted in the MIR images shown throughout the proposal, except for a fourth image we propose to investigate, which provides absolute quantification of refractive index or mass density of the object.
As we shall explain later in the proposal, there are many ways of computing the three MIR parameters from the raw data g(8; x, y), but all involve an implicit or explicit deconvolution (inversion of the imaging model).
B.4 Diffraction-enhanced imaging (DEI) and MIR
DEI and MIR are two generations of essentially the same imaging method, both were proposed by members of our group, and both use exactly the same hardware setup. MIR is the more recent and more general of the two methods, and has been shown to have significant advantages over the original DEI method [4],[9]. DEI can be viewed as a special case of MIR, so we will favor the term MIR throughout the proposal, except when referring to prior results obtained specifically using the earlier DEI method. The principal distinctions between DEI and MIR are: 1) MIR
--- ~ ~

reduces inaccuracies that can occur in DEI images; and 2) MIR produces three images, whereas DEI produces only two (absorption and refraction).
DEI and MIR have shown tremendous promise, and it is now time to begin moving this technology toward clinical application. Our group, which pioneered both methods, has the necessary expertise to make this important next step.
C. PRELIMINARY STUDIES
C.1 Evidence for MIR breast cancer contrast The following is a review of preliminary results which suggest that MIR can produce dramatic increases in image contrast of breast cancer features, as compared with conventional radiography, where many important features can be almost invisible. As explained earlier, DEI and MIR are variations on the same technique, with MIR producing one additional image, and removing some inaccuracies that can occur in DEI.
Therefore, while many of our preliminary results were obtained using the earlier DEI method, these are generally indicative of what can be expected from MIR, and can be regarded as a lower bound on MIR performance.
The following are preliminary results of breast imaging using DEI and MIR.
C.1.1 Visualization of fibrils with DEl Figure 6 (left image) shows a radiograph of a breast tissue specimen that contains invasive lobular carcinoma. This sample has undergone histologic evaluation to confirm that the fibrils in the white box correspond to fingers of tumor extending from the surface of the tumor [24]. DEI images and a radiograph of this region are shown in an expanded view in the right part of the figure. In these expanded views it is clear that tissue contrast is far higher in the DEI images (bottom row) than in a conventional radiograph (top), where the structures of interest are barely visible.
In addition to the usual absorption and refraction images produced by DEI, an image is shown which corresponds to the peak of the measured data function g(B; x, y) . This so-called "peak image" can be thought of as a "poor man's" version of the MIR USAXS image (these studies were performed before MIR had been discovered).
The MIR scatter image measures the width of the scatter function; the peak image observes the USAXS effect indirectly through the relationship between peak value and width of the function. Therefore, this peak image is roughly indicative of the information that will be contained in MIR's USAXS
image.
To quantify the improved contrast of DEI, contrast measures of the fibrils were computed along the image profiles shown as vertical white lines in Figure 6, and this was repeated for other regions of the tissue sample. A
statistical analysis showed that the DEI refraction image had 8-14 times more contrast than a conventional radiograph, while the peak image had 12-33 times more contrast than a radiograph [24].

To determine the potential use of DEI for detection of benign and malignant breast tissue structures, an initial reader study was performed comparing a synchrotron implementation of DEI to digital mammography. This preliminary study compared seven breast tissue samples containing 10 regions of interest verified by histopathologic -- C~-review. The spiculations and architectural distortions present in these regions were determined to represent ductal carcinoma in situ with surrounding fibrosis, infiltrating ductal carcinoma, infiltrating lobular carcinoma, fibrosis alone, biopsy site changes, and fibrocystic changes. The tissues were carefully selected to represent the majority of benign and malignant features commonly seen in breast imaging. A reader study using one expert radiologist demonstrated increased visualization of lesion speculations and architectural distortions in six of seven cases (86%).
This study conclusively demonstrated the ability of DEI to visualize key structural features in breast tissue, but little insight was gained as to what contrast mechanisms of DEI were providing enhanced visualization [12].
C.1.2 Visualization of fibrils with MIR
We have recently begun performing MIR imaging studies of breast tissue specimens. Figure 1 on page 62 shows an infiltrating ductal carcinoma. The branching structures seen in Figure 1 have been interpreted by Dr. Pisano, an expert radiologist, as fibrils extending from the tumor into surrounding tissue. The fibril structure is barely visible in the digital radiograph, but is shown in great detail in all three MIR images, especially in the refraction and USAXS
images. We hypothesize that this additional visual information will help radiologists to discriminate malignant lesions from benign ones.
C.1.3 MIR is not degraded by tissue thickness We are sometimes asked whether MIR can image thick portions of tissue, such as uncompressed breast. The concern expressed is that USAXS might build up to such an extent that it overwhelms the results. While this might occur in imaging of the chest (because of strong USAXS in the lungs), we have found that it is not a problem at all in imaging of the breast or joints (knee and ankle) [4],[5]. For example, Figure 7 shows MIR images of an uncompressed whole-breast specimen, which shows no adverse effects due to tissue thickness; whereas, the radiograph does show the characteristic haze caused by scatter in a large thickness of tissue.
C.1.4 Visualization of calcifications with DE/ and MIR
The features seen in Figure 7 are a collection of benign calcifications. These calcifications are fairly large, so they are visible even in the radiograph, but they are shown in much greater detail in all the MIR images. We have not yet had opportunity to obtain MIR images of a specimen with microcalcifications, however a raw image (Figure 8) acquired in prior DEI research shows what we can expect. This peak image, also obtained in an uncompressed specimen, shows a constellation of microcalcifications as bright points. We hypothesize that the cause of the high contrast of these microcalcifications is a high degree of USAXS, like we see in the benign calcifications in Figure 7.
This is expected, because USAXS measures textural structure on the order of microns, which is expected to be present in calcifications.
~lo-C.2 Preliminary evidence of technical feasibility: hardware issues Using computer-simulation software we have written, there is evidence to suggest that MIR will be feasible using a conventional x-ray tube. In our initial proposal submission, we omitted some of the details of our preliminary feasibility calculation, and the reviewers requested more information on this.
The following section describes the details of one possible imaging system layout that we have simulated. A more complete description of the software used to perform the calculation is postponed to the Methods section.
C.2.1 Preliminary x-ray fluence calculation We have written a software package that uses optical ray tracing to calculate patient dose and track x-ray fluence through an MIR imaging system, based on a specified arrangement and specification of the source, crystals, object, and detector. Because the crystal optics reject x-rays traveling in undesired directions, the main feasibility hurdle for MIR is to obtain sufficient numbers of photons surviving to reach the detector plane.
Using our software, we have done an initial performance calculation for one preliminary design for a compact MIR system. A list of the specifications and results of the simulation are given in Table 1. Schematic and scale drawings of the system layout are shown in Figure 9.

System Parameters In the following sections, we describe the Assumptions imaging system that was simulated, then we Pixel Size 50 m x 50 m describe our findings about achievable fluence and Source to Pre-Mono 15 cm absorbed dose. Source to Object 83 cm X-ray source Source to Detector 100 cm In the preliminary design shown in Figure 9, we Si(4,4,0) Bragg Angle 6.25 degree assumed the source to be a Siemens DURA Akron Min C stal Length 9.2 cm B x-ray tube. This tube, which is used in CT Min Take-Off Angle 1 degree scanners, has a tungsten (W) target, thus it produces Electron Spot Size 12 mm x 0.8 mm Ka,, x-rays at 59.3keV. A powerful tube is needed Est Flux from W Target 435 Ktl, photons/mA-s @ 150 kVp in MIR to achieve the flux needed to overcome Charge Required 120 mA-sec losses in the crystal optics system before the beam Energy Required 18 kJ
strikes the patient. Conventional mammography is typically performed at lower energy (-18-20keV) Current @ 60 kW 400 mA @ 150 kVp than the W source provides, but the refraction and Image Size 20cm x 25cm USAXS contrast used in MIR will persist strongly Imaging Time 6 sec at W energy where absorption and dose to the Results patient are dramatically reduced. The Siemens tube 5cm breast compression has a rotating anode which dissipates heat, and permits the tube to run at high power (60 kW). The Fluence at detector 564 photons/pixel simulated MIR system uses a line-source port on Mean glandular dose 0.004 mGy the tube. 10cm breast compression Optics Fluence at detector 200 photons/pixel This preliminary design has three crystals: a pre- Mean glandular dose 0.012 mGy *
monochromator, a monochromator, and an analyzer. All crystals are silicon, and are tuned for 'Worst-case estimate, which assumes that all attenuation leads to the (4,4,0) reflection order. This crystal order is energy deposition in tissue.
preferred because large crystals can be made by Table 1. Parameters of preliminary MIR imaging system design used slicing along this direction. Such crystals are in computer simulation.
readily available and do not require any new technology to be developed.

~ I I _ Scanning and detectors The scan protocol assumed in this simulation is the same as that used in existing clinical line-scan mammography systems, where experience has shown that motion blur is not a problem if the scan is limited to 6 sec [25]. The detector can either be a single line device that is read out once per image line or an area or full-field device that is scanned in synchrony with the motion of the object across the x-ray beam. In either case, one line or strip of image data is acquired at a time. Standard mammography detectors may not provide optimal x-ray absorption efficiency at W energies because the absorption substrate is thinner than desired. For an eventual clinical system, we envision using direct x-ray-to-charge conversion detectors (see Fig. 10), which allow the use of thick absorbers to achieve efficiency at higher energies without significant loss of spatial resolution. In addition, detector materials with higher Z and density could be employed such as CZT, Pb12 or HgI2 to improve high energy performance. Dr. Yaffe's group has a great deal of expertise in this area, and will provide guidance on appropriate assumptions about detector characteristics that will be used in the proposed simulations.

Results - Fluence Our prelminary simulation showed that fluence at the detector would be around 600 ph/pixel, which is about 1/3 to 1/9 that of a conventional mammogram. Thus, the software predicts that the noise level would be approximately 1.7 to 3 times greater than in a conventional mammogram. . However, our preliminary studies show that, at low noise levels, the refraction contrast can be 8-33 times higher than in a standard mammogram. Therefore, with the added noise, the contrast-to-noise ratio of MIR images would still be about 3-19 times greater than that of a conventional mammogram. Performance may be even better than this if the proposed research leads to improved designs or image-processing methods. Note that, the fluences quoted for MIR are the total for all the raw MIR images combined, therefore these values can be compared directly with conventional mammography.
Results - Absorbed dose For the simulated system, the mean glandular dose would ._.I~_ be 0.004 mGy, which is about 250-750 times lower than in a conventional mammogram at 5cm compression. At 10cm compression, absorbed dose in MIR would be 0.019 mGy, which is thousands of times lower than one would obtain in a conventional mammogram at the same compression. These dose reductions are possible because MIR's refraction and USAXS images are based on refraction contrast rather than absorption contrast, and thus can work at high x-ray energy (60keV) where absorbed dose is low.

C.2.2 Log-spiral focusing element In Section D.3.1, we will describe a new optical component we have developed, a log-spiral element, which may significantly boost the photon throughput of a compact MIR system by creating a bright virtual line source from x-rays emitted by a conventional tube. We have conducted preliminary tests using a bent silicon crystal ((4,0,0) reflection) to focus Cu Ka, radiation from a bent copper plate caused to fluoresce by illuminating it with synchrotron radiation. The x-rays were successfully focused to a line, which was confirmed by recording the intensity at the line focus using an image plate.
Further details of the log-spiral optic will be discussed later in the proposal.
C.2.3 Experience in prototype construction Although no prototypes will be constructed in this project, we do have experience with prototype development, which will guide our proposed feasibility studies. In prior work, we constructed a crude DEI imaging system from scavenged parts, which demonstrated for the first time that DEI images can be obtained using a conventional x-ray tube. This system was not suitable for breast imaging, and is no longer operational. However, the process of constructing the system gave us considerable experience in the practical implementation issues involved, and proved that DEI (and by extension, MIR) can indeed be accomplished in the laboratory using conventional x-ray sources. Figure 11 (left) is a photograph showing the system; Figure 11 (right) shows some initial phantom images produced by this setup.
C.3 Synchrotron-based imaging system Our group has programmatic access to the X-15A beamline at the National Synchrotron Light Source (NSLS) at Brookhaven National Laboratory (BNL). A working MIR imaging system already exists at this facility, and was used to produce all the preliminary results shown in this proposal. Figure 12 is a photograph of the MIR system at BNL. This system offers an excellent testbed for performing the proposed feasibility studies. As described later, images at different noise levels can be emulated by reducing the fluence of the synchrotron beam using absorbers.
C.4 Preliminary theoretical and image processing findings C.4.1 Theoretical noise model We have made some preliminary steps toward understanding noise effects in DEI
and MIR. We recently derived for the first time a theoretical model for the statistics of images computed by the DEI method [26). In this process, we found that: 1) the data must be acquired in a particular way to achieve unbiased results with DEI (which is only possible if there is zero USAXS); 2) both images computed by DEI (absorption and refraction) have signal-dependent noise that is approximately Gaussian; 3) the only imaging-system parameters that influence the noise variance are the beam intensity at the object and the slope of the intrinsic rocking curve; 4) the absorption signal influences the noise level in the refraction image, but not the other way around; 4) the two DEI images are uncorrelated with one another; and 5) the pixel values within each image are statistically independent of one another.
In this project, we will apply the same mathematical formalism to discover the theoretical noise properties of MIR.
C.4.2 Theoretical detectability model We constructed a simple theoretical model to help us understand the relative value of the refraction and absorption signals in the complete absence of USAXS [27]. We chose this simplified case as a basic starting point in developing our analytical approach. We used a simple test statistic to model the visual detectability (in the presence of noise) of a small tissue feature in a simplified analytical phantom (a cylindrical blood vessel in lung tissue; see Figure 13). The evaluation was done as a function of the size of the blood vessel. We found that, for a blood vessel of small diameter, the detectability is much greater in a DEI refraction image than in a DEI absorption image. As the diameter of the blood vessel increases, detectability in the absorption image catches up to that in the refraction images, but by this point the object is clearly visible and detectability is no longer in question.

C.4.3 Initial image processing findings Based on the planar phantom study described earlier, we performed a computer simulation to determine roughly the extent to which MIR images might be degraded by introduction of Poisson noise in the raw data. The purpose of this study was merely to show that MIR imaging algorithms are reasonably robust to high levels of photon noise. The top row of images in Figure 14 show roughly how the images produced by an MIR system might look, at least in terms of noise variance, if the maximum mean photon count in the entire raw data set were only 50 ph/pixel. The MIR images (top row) show that the image quality remains reasonably good, despite the noise. The DEI images in the bottom row show the problems that occur in the original DEI method, which are exacerbated when noise is present, but are resolved by MIR.
Again using computer simulated data, we have also done an initial statistical evaluation of bias and variance in MIR images processed by various noise-smoothing methods. Figure 15 shows plots of standard deviation vs. bias for two small regions of interest (ROI 1 and ROI 2) in a refraction image of a knee joint. Each curve shows the set of possible performance trade-offs that can be obtained by varying the smoothing parameter in a given algorithm. The case of zero smoothing is represented in both plots by the point at which the curves meet the top edge of the graph (here there is high noise variance, and small bias). We compared a Bayesian method (MAP), an Anscombe transform followed by a Wiener filter (AW), and simple 2D lowpass Butterworth - I q -filters applied before (LP-inputs) or after (LP-outputs) the DEI computation.
We found that, using the AW method, it was possible to reduce the noise standard deviation by a factor of about 4-6 while maintaining nearly zero bias.
This suggests that noise in MIR can be effectively combated by appropriate and carefully evaluated image processing algorithms.
C.4.4 Multiple-image computed tomography (MICT) To perform MIR in a computed tomography mode (MICT), it is essential that the three MIR parameters vary linearly with object thickness. This is also a desirable property for image visualization, as a nonlinear image can be difficult to interpret. Under the convolutional imaging model that describes MIR, the three MIR parameter images are indeed linear in object thickness. We have validated the linear model using a series of phantom studies, which we will describe in an upcoming paper. Since our initial submission, we have also worked out a full physical model of MIR, which also validates the linearity assumption. Figure 16 shows an important result, obtained by using a wedge-shaped phantom consisting of microspheres suspended in glycerin, which shows that USAXS is a linear function of object thickness. A least-squares fit showed the relationship to be linear at a significance level ofp < 10"4.
Next, we show preliminary MIR tomography (MICT) experiments, in which we have imaged a physical phantom and specimens of cadaveric bone using the synchrotron. A scliematic of the phantom is shown in Figure 17; a reconstructed 2D slice image of the phantom is shown in Figure 18.
Figure 19 shows MICT images of a bone specimen. These tomographic images were computed by applying ordinary filtered backprojection (FBP) to planar MIR images. The results demonstrate that CT is indeed feasible; however, further research will be needed to develop and validate MICT.

D. RESEARCH DESIGN AND METHODS
D.1 Overview of research plan The overall goal of this project will be to predict the best image quality that might be obtained by a future clinical MIR device. We will then use a reader study to evaluate whether such images will demonstrate clinical information beyond that provided by conventional mammograms. This will determine whether a clinical MIR system can be feasible, and whether building such a device in the future would be worthwhile.
To accomplish our goal, we will use computer simulations, image-processing research, and benchtop measurements to develop and simulate a set of optimized, candidate designs for a future clinical MIR imager.
Candidate designs will include options ranging from using off-the-shelf components only, to designs using more costly and specialized components. The final reader studies will tell us how complex and costly a future system must be to achieve clinical effectiveness, and whether a compact system can significantly outperform conventional digital radiography. For example, while a prohibitively expensive table-top synchrotron might satisfy the design requirements, and might one day be a viable option, we are hopeful that a much more practical approach will work as well.

-- ~5 -Figure 20 illustrates the conceptual MIR block diagram that we propose to optimize, so as to predict the best achievable performance characteristics. Although the three steps in this chain-data acquisition, hardware design, and image processing-are all interconnected, much of the work to optimize each part can be performed separately.
For instance, our analyses have shown [26] that the image-processing step depends essentially on only three factors in the hardware design: 1) x-ray energy of the source (which influences the MIR parameters of the object); 2) output x-ray fluence (which determines the noise level); and 3) the shape of the intrinsic rocking curve (which determines the convolutional imaging model). Therefore, the image-processing team will aim to produce plots of performance vs. x-ray energy, noise level, and rocking curve type ((1,1,1), (3,3,3), etc) for image features relevant to breast cancer. Fluence predictions from the hardware research will tell us which points on these curves can be achieved by practical hardware strategies. With the exception of x-ray energy, the choices of data-acquisition parameters do not influence the hardware design at all, but do influence the image processing.
Therefore, optimization of data-acquisition parameters and image processing methodology will be treated as a co-design process.
Having introduced our overall strategy, let us briefly outline the proposed research. Aim 1 involves development of a basic understanding of MIR contrast in breast imaging, leading to development of an atlas of quantified and pathology-proven cases, and construction of physical phantoms.
Aim 2 consists of optimization of the hardware design, image processing methodology, and data-acquisition parameters. Aim 3 is a reader study to determine the diagnostic utility that can be expected of a compact clinical MIR imager. Aim 4, which is a byproduct of the design optimization in Aim 2, is to compile specifications of candidate designs for the first clinical MIR
prototype, which would be constructed under a future project if feasibility is established in this project.
Having introduced the overall plan, we next describe the specific steps we will take to accomplish each of the specific aims.
D.2 Aim 1: Understand MIR contrast mechanisms in breast imaging, and construct realistic MIR breast phantoms In this portion of the work, using representative specimens of breast tissue, we will work to develop an understanding of MIR contrast mechanisms as they relate to breast imaging, and develop phantoms to guide the design optimization in Aim 2. Aim 1 will proceed in three steps: 1) image breast specimens at best sychrotron quality, then perform histopathologic correlation; 2) based on the findings of Step 1, determine MIR parameters of important features in breast tissue; and 3) based on the findings in Step 2, construct realistic breast phantoms, designed to replicate the three physical effects seen in MIR, for use in Aim 2.
D.2.1 Aim 1, Step 1: Correlation of MIR image features with pathology findings We will begin our study of MIR and its depiction of breast cancer features by imaging, at best synchrotron quality, 24 mastectomy specimens from UNC Hospitals. These images will guide the initial phases of the research, providing us with relatively noise-free indications of MIR image features.
Specimens will be selected to include: six cancerous masses, six noncancerous masses, six cancerous clusters of calcifications and six non-cancerous clusters of calcifications. Tissues selected for the study will also represent a range of tissue thicknesses, from a 1.0-cm surgical pathology specimen to 4.5-cm or greater mastectomy specimens.
Thin specimens will allow for a detailed characterization of tissue structures, while thick specimens will provide insight into how well MIR will perform at realistic tissue thicknesses.
This study seeks to characterize both malignant and benign features present in breast tissue, which will provide insight into both the sensitivity and specificity of MIR. Previous studies of DEI in breast tissue have emphasized visualization of malignant structures, such as invasive lobular carcinoma. While this is important, the real diagnostic utility of MIR will lie in its ability to distinguish between malignant and benign structures, which is why we will carefully study both malignant and benign cases.
Subsequent to imaging of the specimens at the NSLS at BNL, all tissues will be sectioned and mounted for histopathologic correlation. Dr. Chad Livasy, M.D., UNC Department of Surgical Pathology, and Dr. Etta Pisano, M.D., UNC Department of Radiology, will relate the features present in the MIR
images to the physical and pathologic structures present in the actual breast tissue.
For accurate correlation, images will be presented on a high-resolution, high-brightness monitor situated directly beside a microscope for histologic evaluation and interpretation.
Every region of interest visualized in the MIR image will be compared with a conventional digital radiograph, and located in the actual tissues. In contrast to conventional mammography, no correlation studies have ever been performed to determine how breast tissue - /6_ structures and pathologic lesions appear using MIR. Mapping visual findings to the actual structural and pathologic features is critical to establishing the validity of MIR for diagnostic imaging.
D.2.2 Aim 1, Step 2: Determine characteristic MIR signatures of tissue features Careful MIR imaging of the specimens in Step 1 will allow us to determine precisely the absorption, refraction, and USAXS properties of key breast-tissue features. As we will explain shortly, for purposes of this portion of the research, we will convert the refraction-angle images r(x, y) to images depicting absolute quantification of refractive index n(x, y) or mass density p(x,y) (mass per unit volume). The reason for this step is that, while the refraction-angle image is excellent for visualization, it is not a pure material property, because refraction angle is a function of both refractive index and object shape. The mass density, on the other hand, is a more useful quantitative measure of tissue, and also will help guide the identification of suitable materials for the phantoms.
For each tissue feature identified by Drs. Pisano and Livasy, we will compute its MIR signature, i.e., a vector of parameters x = (a, p, s), describing its absorption coefficient, mass density, and USAXS parameter (a measure of subpixel tissue texture). Each image feature will be segmented by hand, and the mean of its characteristic x vector will be computed. We will also employ simple unsupervised clustering algorithms, such as Gaussian mixture models [28], to identify any distinct portions of each tissue feature. By this process, we will compile a visual and quantitative atlas of the appearance of many major tissue features seen in breast imaging with MIR.
A side benefit of this work will be early indications as to whether the x-vector signature of a tissue feature, such as a calcification, is sufficiently informative to allow tissue features to be automatically classified by computer.
This will give us an indication of whether MIR pixel classification may be a helpful tool, in the future, for computer-aided diagnosis and assistive visualization methods.
Refractive-index/Mass-density calculation Conversion of the refraction-angle image r(x, y) to a refractive-index image n(x, y) appears to be feasible, but some additional research will be needed to perfect this calculation. The refraction-angle image r can be expressed as the derivative of the projected refractive index n of the object, i.e., r(x, y) = a10'y jn(x, y, z)dz , which can be alternatively expressed in terms of the mass density p of the material by using the fact that n- xp [32], where the coordinates (x, y, z) are as shown in Figure 3 and K is a physical constant proportional to the x-ray wavelength squared. Therefore, an image of refractive index n(x, y) tells us about the projected mass density of the tissue at every pixel.
It would appear that the refractive index can be easily obtained by inverting the expression for r(x, y) by integration. However, in practice, this method results in unusable images, as shown in Figure 21 (lower left). Even a small amount of noise in the data can cause severe streaks that run across the length of the image in the y-direction.
In initial studies, we have found that a linear minimum mean-square error (LMMSE) technique can perform reasonably well in the presence of noise (Figure 21, lower right); however, some refinements will still be needed.
We will try two additional methods-projections onto convex sets (POCS) [29]
and maximum a posteriori (MAP) estimation-to determine the best results that can be obtained, and evaluate the methods by using real-data images of objects of known refractive index, including spheres of appropriate materials (see next section).
D.2.3 Aim 1, Step 3: Use findings of Step 2 to construct realistic breast phantom for MIR
Having identified the true (a, p, s) values of tissue features proven by histologic examination, we will be in good position to construct a set of realistic breast phantoms, which will convey all the key properties needed for MIR.
Note that standard mammography phantoms are unsuitable for MIR, because they aim only to simulate gross x-ray absorption. Our team members at Sunnybrook & Women's, led by Martin Yaffe, have significant expertise in ~I -~-designing phantoms both for magnetic-resonance imaging and mammography;
therefore, we are confident that we can produce an MIR phantom that captures all three of the effects measured by MIR.
A set of phantoms for evaluating MIR will be developed with the aim of producing the following characteristics: 1) the phantoms will have structures that mimic those found in the breast, including normal tissue, masses, fibrils, and calcifications; 2) the materials used will demonstrate similar x-ray absorption, refraction and USAXS characteristics to those in the breast; 3) the objects in the phantom will have similar sizes to those found in the breast; 4) to achieve appropriate USAXS, the materials will have similar textures on a scale of 10-40 m to those found in the breast; and 5) the background of the phantom will behave similarly to that of normal breast tissue. The process of phantom development will be an iterative one (repeated construction and validation); but will proceed essentially in the following phases.
The first phase will involve identifying materials that have refractive indices and USAXS properties that are similar to biological materials. This will involve testing a number of prospective materials. For the background of the phantom, which should provide a semi-rigid support matrix, we will try a number of tissue-equivalent materials.
Sunnybrook has developed a number of materials which are tissue-equivalent for both radiography and magnetic-resonance imaging. These are uniform materials with various agar-gelatine-oil ratios which mimic the attenuation of fatty or glandular tissue, but have no internal structure. A number of commercial filaments (nylon, polyester and silica) will be placed in the background material and tested to see which most closely mimic fibrils, and whether fibrils are more correctly modeled as simple fibers, or as fiber-shaped objects containing texture that produces USAXS. For calcifications, calcium hydroxyl-apatite which is known to mimic microcalcifications in conventional radiography will be tested. Masses will be represented by embedded objects cast from glandular-equivalent material. A number of samples of microspheres of different sizes and compositions will be evaluated. Materials which show the required effects will be characterized more fully and graded in size.
After identifying the best materials, phantoms will be constructed that will allow us to test the sizes of objects which are visible, and measure signal dependence based on the size of the simulated fibril or calcification.
The size will be reduced by factors of 1/4 until the object is about one-tenth the size of a pixel. Another set of tests will evaluate the effect of multiple fibrils and calcifications, and the effect of inter-object spacings. This will involve developing methods to accurately space fibers and calcium specks in the agar background.
Another set of phantoms will test the sensitivity of the MIR image to the angular orientation of the fibrils, the sensitivity to the depth of the fibril within the object and the effects of having randomly distributed fibrils.
Additional tests will be made of directionality, and packing density effects, as well as the effect of stacking fibers.
A final breast phantom will be developed, which has a mildly cluttered background to represent tissue in the "average" breast, and has masses, fibrils, and microcalcifications of various sizes for use in the reader detectability studies and other preliminary image evaluations to be performed in Aim 2b.
D.3 Aim 2 - Optimize MIR imaging performance under constraints for a compact clinical device As explained earlier, the principal challenge in developing a clinical MI.R
device is to avoid or, at least, minimize image degradation due to photon noise. In Aim 2, we will work to meet this challenge by optimizing the hardware design, image-processing methods, and data-acquisition parameters to predict the best performance that might be achieved in a practical future imaging system. Aim 2a will focus on optimization of the hardware; Aim 2b will focus on optimization of image processing and data acquisition, performing quantitative image-quality assessments, and determining the best overall configurations/images for the final expert-reader study.
D.3.9 Aim 2a - Hardware optimization As in any complex optical system, relationships among the parameters of an MIR
imaging setup are highly interdependent, and impossible to predict without computer-aided design software. Therefore, we will develop design software based on optical ray-tracing, validated component specifications, and standard radiometric calculations. We will then use this software to develop four candidate designs for a compact imager.
We have made a significant start on the design software, which was used to generate the rough prediction of x-ray fluence described in the Preliminary Results section. However, many aspects of the software need to be upgraded to reflect hardware characteristics measured in actual laboratory experiments. Thus, laboratory experiments will be conducted to establish the correct assumptions to be made in the software, and to validate the output results. These experiments will be performed using a conventional laboratory setup at the University of Saskatchewan and also the synchrotron at BNL. Based on user-defined input parameters of the imaging system, the _( R_ software will compute the output fluence and intrinsic rocking curve of the system, and will check to ensure that the design parameters remain within input specifications.
The results for photon flux and intrinsic rocking curves, obtained in the hardware design studies, will be used by the image processing team to determine achievable image quality of the system. In the final year of the project, the best hardware designs, coupled with the best corresponding image-processing methods and data-acquisition parameters, will be used to guide final preparation of images for an expert-reader study to determine clinical performance of MIR.
The steps of the hardware optimization research will be as follows: 1) perform laboratory experiments to determine performance specifications of key components; 2) redesign our existing modeling software to incorporate the new components, imaging situations, and parameter assumptions; and 3) use the finished software to produce four design options for the first prototype MIR breast imager. These steps are described in the following sections.
Aim 2a, Step 1: Laboratory experiments for hardware characterization To obtain the information needed to upgrade the design software, we will conduct three sets of laboratory experiments: 1) testing of a new log-spiral focusing element designed to improve output fluence of the system; 2) measurement of fluence throughput of a conventional source coupled with slightly mismatched pre-monochromator/monochromator crystal pair; and 3) flux measurements, with the crystal optics in place, as a function of accelerating voltage in the x-ray tube. These three sets of experiments, which will provide the information needed to complete the hardware-design software, are explained in the following sections.
Experiment 1: Log-spiral focusing element In this experiment, we will test a new optical component that promises to boost photon flux significantly in a conventional-source MIR imager. The optical component will be constructed and tested in the laboratory, and then its properties will incorporated in the design software, so that it can be included in the design studies.
Recall that in MIR, we ideally require an x-ray source that is very intense, so that sufficient x-rays reach the detector to form a good image;
but the source must also be reasonably small, so that system resolution is not degraded by source-size effects. Unfortunately, it is difficult to make a conventional x-ray source that is both small and intense, because of the problems associated with dissipating heat -generated by an intense electron beam impinging on a small target area.
One of us (Chapman) has developed a potential solution to this problem [30], which may significantly boost the output fluence of an MIR imager. The solution is to use an x-ray source with a large target area, which can achieve high power, and then to focus the emitted radiation to form a thin, virtual line source (see Figure 22). The virtual line source will be small in the dimension that matters, and very bright; therefore, it is expected to have the properties needed to reduce, or even overcome, the problem of inadequate fluence at the detector.
The focusing element, which we call a log-spiral, is a bent diffraction crystal, the surface of which is a portion of a logarithmic spiral. Mathematically, this is the surface shape required to make the Bragg-diffracting element behave as a focusing device (see Figure 23). The log-spiral element has the following properties: 1) it collects light emitted from a large target area at a fixed take-off angle where brightness is at a maximum; 2) it monochromates the beam; and 3) it focuses the radiation to form a high-brightness, virtual line source, as shown in Figure 23. The small width of the line source will produce good spatial resolution, which can be improved even further by placing a slit at the focus.

1q In this project, we propose to conduct the first experimental validation and characterization of the log-spiral element. We will construct the laboratory-based characterization system shown in Figure 24. Silicon (3,3,3) planes will be used for the log-spiral optic, due to the ready availability of inexpensive silicon wafers with this orientation that can be bent to the required shape. The research plan will proceed as follows. 1) We will model and measure the reflectivity of bent Bragg crystals using bender frames of fixed radii. These measurements will be done using a conventional x-ray source. 2) Based on optimal reflectivity for reasonable bending radii, a bender frame will be fabricated for the initial tests. 3) Using the TIGER code (developed at Los Alamos for x-ray flux yields from electron impact on targets), we will optimize the x-ray flux from an anode under conditions allowed by the x-ray power supply and acceptable power loading conditions. The optimization parameter is the electron impact energy and angle to the anode and the photon take-off angle. 4) We will design and model the appropriate caustic spiral anode for the tube source based on the optimized photon flux take-off angle and electron impact angle and energy. 5) We will measure the flux from the anode (shaped target) that can be collected by the log spiral optic under ideal collection conditions and then input into the monochromator where flux will be measured under a variety of conditions (beam size, focal distances variations, flux through apertures at the focal line, etc.).
Experiment 2: Conventional source flux through pre-mono/mono pair The flux throughput will be measured for a conventional x-ray source, when combined with a pre-monochromator and monochromator (a slightly dispersive mismatched system; i.e. Germanium (3,3,3) and Silicon (3,3,3) to prevent unwanted photon energies from traversing the imaging system). These measurements will consist of the flux into a specified area that is the same size or smaller than the beam prepared by the optics. This will determine the flux that can be accepted by the optics. In addition, the effective source size of the system will be evaluated, using small apertures, to ensure that sufficient spatial resolution is obtained. These measurements will be made using scintillation detectors, an x-ray tube, crystal optics, and two sets of slits.
We will assess the brightness of the beam (flux emitted per source size per solid angle), and the transverse acceptance of the mismatched crystal system. This acceptance will be explored with different mismatched crystal types and an assessment of crystal availability and crystal quality will be made.
Experiment 3: Flux throughput as a function of accelerating voltage The flux measurement will then be made as a function of accelerating voltage.
It is well known that the flux from an x-ray tube emission line is a super-linear function of accelerating voltage to the point where the energy of the electrons is sufficient to penetrate the anode material significantly and/or not interact effectively with the core electrons. For tungsten, the optimal accelerating voltage is approximately 150-200 kVp. Fluence measurements will be made to the safe limit of an existing x-ray tube unit and extrapolated to higher x-ray energies using the Integrated TIGER Series (ITS) computer code [31,32]. These fluence measurements will be made with a dispersion mismatched system.
Aim 2a, Step 2: Upgrade of hardware-design software In this portion of the research, we will make essential refinements to the present version of the hardware-design software we have developed, to ensure that we can accurately predict MIR
hardware performance with conventional x-ray tubes and, optionally, with the log-spiral optic in place.

Based on more than 20 user-specified input variables the current software produces output brightness values, predicted fluence at the detector, and schematic drawings of the system layout. Currently, the input parameters include quantities such as desired spatial resolution, properties and placement of the crystal optics, size and composition of the object, and specifications of the x-ray source and the detector.
The software enhancements that will be made in the project are as follows.
1) Performance calculations based on various x-ray tube types, which are currently based on extrapolations from Mo performance, will be modified so that they are based on measured values.
Specifically, we will include tungsten and silver flux values, as a function of accelerating voltage, measured with the crystal optics in place. These measurements will be made in Step I of the hardware-design work, as described earlier.

--ao--2) The program will be modified to include a synchrotron source as an option.
The synchrotron option will serve as a tool for validating the program, since we have a synchrotron-based system at our disposal. In addition, this will allow us to predict performance from table-top synchrotrons that are currently under development.
3) The system arrangement calculations are currently done independently in two orthogonal planes that track the central-ray beam trajectory, which introduces a discrepancy in distances. A
full three-dimensional arrangement will be implemented which can be imported into a 3D visualization program such as AutoCAD.
4) The program will be upgraded to allow for investigation of designs that incorporate the log-spiral optic described earlier, including the optical modifications and flux limitations this component will introduce.

5) The program will be modified to produce simulated MIR images, by coupling this code to existing physics/noise modeling code that will be provided by the IIT team. When combined, the software will produce simulated MIR
images for a phantom object under user-defined imaging conditions and hardware configuration.

6) In the current software version, calculations are based on achieving the same fluence as in a Mo based system.
The calculation will be changed to reflect the required fluence to obtain a specific image quality (e.g., signal-to-noise ratio) for the absorption image, refraction image, USAXS image or some combination within defined limits.

7) The program will undergo a rigorous testing phase to assess the accuracy of the outputs. The calculations will be matched against measured values in existing systems, both laboratory-based and synchrotron-based.

8) A more friendly user interface will be added so that the software will be accessible to a wider group of users.
Aim 2a, Step 3: Use design software to develop candidate hardware designs In this portion of the research, we will work to develop four hardware design options for a clinical MIR imager.
These design options will be used in Aim 2b to predict the best image quality that a clinical MIR device might produce in the future. For each design option, using computer-simulation software, we will aim to maximize the fluence at the detector (and, thus, minimize photon noise) within practical specifications. The designs we will develop will include ones based only on commercial off-the-shelf (COTS) components, as well as more unusual designs that include custom-made components. Our four candidate designs will Source be based on combinations of the following design choices, as shown in Table 2.
COTS Optimized Commercial off-the-shelf components (COTS). For the source, COTS refers to energy the source being made from a standard target material. For the optics, COTS

refers to standard, readily available crystals. ~
Log-Optimized x-ray energy. In optimized-energy designs, we will allow x-ray O pi al 3 4 energies for which tubes are not normally sold commercially. For example, a samarium (Sm) target produces x-rays at 40keV, and table-top synchrotrons, Table 2 Four design options for which allow for a continuum of energy choices. feasibility study of clinical MIR imager.
Log-spiral optic. These designs will incorporate a log-spiral focusing element.
The design approach will consist of an iterative process in which the design software will be used to develop a specific hardware configuration as follows.
Step 1. A design will be prepared based on the constraints defined by one of the options in Table 2.
Step 2. This design will be reviewed by the local design team to ensure that all specifications are met or met as close as reasonably achievable.
Step 3. The team will incorporate any modifications that will simplify the design or allow specifications to be met or exceeded without violating other specifications. If modifications are made then the next step is Step 2.
Step 4. Once a system based on these criteria has reached this point, then the design will be prepared as a Conceptual Design Report (CDR), and distributed to the larger group for comments. A meeting of the group will be held either in person and/or via web conference to discuss the merits and shortcomings of the design. If modifications are suggested, then back to Step 1.
Step 5. A Final Design Report (FDR) will be prepared which will consist of a set of performance specifications, source specifications, detector specifications, x-ray optical specifications, projected cost, estimated final design and construction schedule.
The Final Design Report for each of the four designs will be a reportable outcome of this project, which could be used in future prototype development work.
D.3.2 Aim 2b - Optimization of image-processing algorithms and data-acquisition parameters Like x-ray CT, MIR is intrinsically an indirect or computed imaging method, meaning that the images are created through a mathematical image-processing calculation. Therefore, the image-processing algorithms are central to the ___ DO-quality of the final results. This will be especially true as we attempt to move MIR from the synchrotron toward a compact clinical imager, where photon noise may become a significant factor.
In this portion of the research, we will also study the influence of the data-acquisition parameters, in particular: 1) the x-ray energy; and 2) the number and selection of the raw images to be acquired. The issue of which raw images should be acquired determines how the hardware is to be used, and not its essential design. Therefore, a study of these parameters is grouped as part of the image-processing research.
The overall objective of this research will be, through optimization of the image-processing methods and data-acquisition parameters, to produce the best MIR images we can from noisy data, since noise is the main potential limitation of a clinical MIR imager. Our group has a strong background in image reconstruction for PET
and SPECT, where Poisson noise is the limiting factor; therefore we are in a good position to tackle the problem of noise in MIR, which will be similar, but less severe.
As a basis for this research, we will use an exhaustive data set acquired using the synchrotron, as described below. As the project progresses, the work done in Aim 2a will tell us the level of noise we can expect from each of the most promising hardware designs.
To guide the research, image evaluations will be performed on an ongoing basis in order to identify the best algorithms and parameter values. These intermediate evaluations will be based on: 1) regular feedback from Dr.
Pisano, a radiologist who is an expert in breast imaging; 2) standard quantitative image-quality measures, including numerical-observer detectability measures; and 3) periodic reader studies involving the scientific staff. The specific evaluation strategies are discussed at the end of this section, since they pertain to all the methods and parameters that will be studied. Once we have decided on the best candidate techniques, we will use the methods we have developed to prepare MIR images for the final reader study in which we will obtain a rigorous evaluation of the diagnostic information provided by MIR.
Acquire exhaustive set of MIR data of the phantom The synchrotron at Brookhaven National Laboratory provides an excellent tool for emulating the kinds of images we might expect from a future clinical MIR imager. The x-ray energy of the synchrotron beam can be varied as desired, and the noise level of MIR images produced at the synchrotron can be varied by inserting absorbers in the beam.
Therefore, the synchrotron allows us to emulate the whole gamut of possible MIR images, making it an ideal and physically realistic simulator of a future clinical device.
With this idea in mind, we will use the synchrotron at Brookhaven to image the phantom developed in Aim 1 at a wide range of x-ray energies, noise levels, and analyzer positions. Once compiled, this exhaustive data set will provide us with the information needed for studies to fully explore parameter space. For example, by finely sampling at 256 analyzer positions, we can study the effect of number and placement of these samples by selecting subsets of samples from the exhaustive set. In essence, the exhaustive data set will provide us with a look-up table of possible data examples for use in our research activities.
We will initially use the synchrotron to image the phantom at four x-ray energies in the range where MIR is expected to operate best (30, 40, 50, and 60keV); 256 analyzer positions (equally spaced from -20 to 20 microradians); and nine noise levels (in octaves from 16 to 2048 photons/pixel, as measured in the background part of the phantom). We will acquire 50 noise realizations for each combination of parameters, for use in validating our noise models.
The range of noise levels chosen spans from highly photon-limited, up to photon-count levels present in conventional mammograms. Noise level will be determined by fixing the integration time to achieve a total breast scanning time of 6sec, consistent with existing clinical devices, and place thick Lucite sheets in the beam to reduce fluence to the desired level. We will mount a step-wedge Lucite absorber on a translation stage, so that most of the image acquisition can be fully automated without opening the imaging hutch at the synchrotron. Lucite is known from our prior studies to introduce no USAXS, and flat Lucite sheets have proven to introduce no significant refraction. This is a standard method of adjusting beam intensity at a synchrotron facility. Should we discover later in the research that any parameter combinations have been overlooked, we will return to Brookhaven to acquire additional data.
Modify MIR algorithms to account for noise For the most part, DEI/MIR literature has ignored noise effects, probably because noise is insignificant in existing synchrotron implementations. In moving to a clinical device, we will need to consider the effect of noise, and build - aa -mechanisms to combat it. Some of the algorithms that have been proposed in the past, including the original DEI
method, do not perform well when noise is present in the data.
Note that, in the following discussion, we will switch to discrete representations of the main quantities of interest, consistent with digital imaging. Thus, we replace the continuous quantities f(9; x, y) , g(8; x, y) , and R(O) with the discrete quantities fm,n [k] , gn [k] , and R[k], respectively, in which m=1, ..., M and n=1, ..., N
are pixel indices, and k=1, ..., K is the index of the angular positions of the analyzer at which the data are collected.
Goal of MIR image processing The goal of MIR processing is, given knowledge of the observed data gm n[k]
and the intrinsic rocking curve of the imaging system R[k], to determine the three MIR images: absorption am n, refraction angle rm n, and USAXS
Sm n The absorption image is defined as the estimated absorption coefficient at pixel (m, n) , defined as am n=-ln(Im n), where Im,n Ek fm,n [k] is the total transmitted intensity.
Refraction is the net average deflection angle of the beam, defined as rm n= centroid ( fm n[k]) - centroid (R[k]) .
The USAXS image is the second central moment (variance) of the object function, defined as Sm,n = Var( fm,n [k] /
Im,n ). The following are some of the alternative strategies we will investigate to optimize MIR image processing.
Deconvolution approach Following a photon-limited noise assumption, our observation model will be as follows:
gm,n [k] - Poisson(R[k] * fm,n [k]) .(In situations where quantum background noise may be significant, this model will be modified to include a bias term in the expression for the Poisson mean.) The aim of the following methods is to invert this model to estimate fm,n [k] ; the differences among the methods lie in the way in which they aim to lessen the effect of noise. To simplify notation, we will sometimes represent images as vectors (indicated by bold letters). For example, fm n will denote the discrete angular intensity spectrum at an individual pixel (m, n), and f will denote a one-dimensional (vector) representation of the three-dimensional function fm n[k] obtained by lexicographic ordering of the pixel values.
Maximum-likelihood solution. The maximum-likelihood (ML) solution of the deconvolution problem is the value of fm n that maximizes the Poisson likelihood function for each pixel. A
constrained solution can be obtained by using the well-known expectation-maximization (EM) algorithm [33]. The EM algorithm for this deconvolution problem produces a solution that is guaranteed to be non-negative if the initial estimate is non-negative. The iteration, which is widely used in nuclear-medicine imaging [34], is described by the following expression:
c> 1 'k k fmtnl)[k]_ fm,n[ ] {R[k]*[R[J)[k]j} [
(1) R[l] 1=1 where fõ,,), [k] denotes the ith estimate of fm n[k] , and * denotes cross correlation. Note that the iterative deconvolution in Eq. (1) is performed separately at each pixel because the likelihood functions are not coupled in any way. In practice, if the noise level is high, this algorithm is usually stopped prematurely to stabilize the solution in the presence of noise. Early stopping of the EM algorithm amounts to an implicit method of regularization.
Based on early success of pre-reconstruction smoothing (see Figure 15), we will evaluate the Anscombe-Wiener (AW) and low-pass filter methods as well as early stopping, to combat noise effects.
Maximum a posteriori solutions. When there is an appreciable amount of noise in the data, an explicit regularization approach may be preferred. We will employ a Bayesian approach based on the maximum a posteriori (MAP) criterion, i.e., f= arg max p(g I f)p(f ), where p(f) represents a prior probability law on f that reflects known properties of the actual signal. In the present context, the prior informs the algorithm that we expect f(B; x, y) to be a smooth function.

--~ 3 -We will represent this knowledge by way of a Gibbs prior, which has the following form:
p(f) = exp[-,QJ,V,. (f )] / Z in which the potential functions Vk (f ) are defined so that smoother functions are favored over less-smooth ones, 6 controls the strength of the prior, and Z is a normalizing constant. A modified EM algorithm [28] can be used to find the MAP solution. The general form of the iterative procedure is similar but involves a term in the denominator that depends on the specific choice of the prior. We will study two ways to define the prior, and thus two specific choices for the derivative term.
Specifically, we can choose to impose smoothness only along the k -axis of fm n[k] , thus encouraging each pixel's angular intensity spectrum to be a smooth function. Alternatively, we can impose smoothness spatially as well, thus utilizing the knowledge that, because of their close proximity, neighboring pixels will tend to have similar angular intensity spectra. Following nomenclature from the image-processing field [31], we describe each pixel's angular intensity spectrum as a channel; thus, we describe the former approach as a single-channel method, and the latter as a multichannel method.
Our research team has used multichannel techniques successfully in other NIH-sponsored research on reconstruction of gated and dynamic nuclear medicine images, and the P.I. has co-authored a book chapter on the subject [35]. The multichannel method appeared to produce the best MIR results at significant noise levels in our preliminary work [4], but the methods have yet to be compared quantitatively.
Model-based approach In the deconvolution approach, we deconvolve g to obtain f from which we measure the MIR parameters. In the model-based approach, which we first studied in [3], we will aim to estimate the MIR parameters directly from g, by using a Gaussian model for the angular spectrum at each pixel in f Like the deconvolution approach, the model-based approach can be performed either pixel by pixel, or as a multichannel method (which is likely to be more successful at high noise levels).
We will explore the following model for g, which assumes that f is a Gaussian-shaped function in terms of its angular coordinate: gm õ[k] =1o exp(-am õ) exp [(9k - rm,,, )2 / 2sm,n R(6k ), where Bk is the angular analyzer position (in microradians) of angular sample k, and Io is the initial beam intensity. A maximum-likelihood solution will be determined under a Poisson likelihood function for g. A multichannel version of this approach will also be developed by solving for all the pixels simultaneously under a Gibbs prior which expresses signal correlations among neighboring pixels. In principle, the model-based approach may prove to be more powerful than the deconvolution-based approach for three reasons: 1) it substantially reduces the dimensionality of the statistical estimation problem, 2) it imposes implicit regularization by assuming a smooth functional form, and 3). if the model on which the method is based is known to be correct, the model parameters should be optimally informative measures of the object function. However, the relative merits of these nlethods will depend on the accuracy of our model, which will be determined by evaluation studies.
Non-Poisson models At lesser noise levels, where the Poisson law can be approximated by a Gaussian distribution, or in cases where quantum background noise is estimated and corrected using a calibration scan, preconditioned conjugate gradient methods can be used, in conjunction with optimized penalized weighted least squares objectives [34]. These approaches offer some benefits in terms of algorithm convergence when Bayesian priors are used [34], and will be explored as part of the image-processing research.
Computed tomography algorithms Upon the successful development of planar-mode MIR, we will extend MIR to a multiple-image computed tomography mode (MICT). This will allow us to measure the properties of a volumetric object on a voxel-by-voxel basis, which may be more informative than planar imaging, especially for tissue characterization.
We will reserve final radiologist reader studies of MICT to a future project;
however, we plan to develop the basic tools so that we will be in a position to plan such studies as part of prototype-development work that would take place under a future project.
In MICT, we will use the scanning geometry illustrated in Figure 25. To acquire data for tomographic reconstruction, the object will be scanned throu hg multiple angular views and in the v-direction as well to cover the --:;), 4 -whole breast. Including the analyzer position 0 and the view angle a, MICT
projection data are four-dimensional, i.e., g = g(B, a; x, y) . Our aim is to compute multiple 3D images from this 4D data set, including the absorption, refraction-angle, refractive-index, and USAXS images discussed previously.
Note that the reconstructed absorption image will be similar to a conventional x-ray CT image, except that it will have much higher quality due to scatter rejection. However, the reconstructed refraction and USAXS images will be entirely new to clinical mammography.
MICT reconstruction can be accomplished in two basic ways, both of which will be investigated. First, we will pursue the simple approach used in our preliminary studies, in which we first compute MIR projection images, then apply standard reconstruction methods to compute the MICT images. Second, noting that multiple 3D images can also be viewed as a single 4D image, we will investigate fully 4D
reconstruction of the entire 4D MICT image from all the raw data, similar to our work in NIH-sponsored projects on reconstruction of brain and cardiac image sequences (e.g., [36]-[39]). These methods will use maximum-likelihood and Bayesian reconstruction methods based on Poisson noise models, which have been well-developed in nuclear medicine applications. To validate the quantitative accuracy of the MICT results, we will use the physical phantoms from Aim 1 and the evaluation approaches described later in this section.
Optimize data-acquisition parameters All of the following parameters will be optimized by using the image-quality evaluation approaches described later in this section.
Energy. The absorption coefficient and refractive index of a material are dependent on the energy of the x-ray beam.
Therefore, selection of the x-ray energy will influence the three MIR
parameters for a given object. Energy will be coarsely optimized by evaluating image quality at the four energies acquired in the exhaustive data set described earlier. If the resulting curves of image feature detectability vs. energy suggest that an intermediate energy may represent the optimal point, we will perform an additional imaging session to acquire phantom data for an intermediate energy.
Number and arrangement of analyzer positions. In MIR, the analyzer crystal is rotated, and raw images are obtained at discrete positions along the way. The analyzer positions at which the raw images are acquired ( Bk , k = I, ..., K) are samples of the function g(B; x, y), which yield the discrete data representation g,n n[k] . An important factor influencing MIR image quality is the number and placement of these angular samples.
We will begin this investigation by identifying the effective angular support (angular range) of the function g(B; x, y) using the exhaustive data set described earlier. Next, assuming a uniform sampling strategy in 0, we will perform image-quality evaluations as a function of the number of sample points (raw images) acquired. This will be achieved by downsampling the exhaustive data set. We will also evaluate image quality as a function of the assumed size of the effective support, so as to minimize the angular range of the measurements as much as possible without harming image quality. This is important so that exposure time is not wasted in collection of raw images at analyzer positions where there is little useful information.
Having identified the minimum necessary angular range, and minimum necessary number of samples, we will next study whether performance can be improved by non-uniform sampling.
Since few pixels will have significant values of g(e; x, y) at the extremes of the angular range, it seems likely that it would be acceptable to use less sample points at these extremes, and thus allocate exposure-time budget principally to angles near the origin, where most pixels will have significant values.
We will simplify the search for the optimum non-uniform sampling pattern by parameterizing the sample spacing as follows. We will use a sampling pattern in which the local sample spacing at position 0 is directly proportional to 10 Id , and study image quality as a function of the parameter d. Thus, the sampling pattern will be constrained to be symmetrical about the origin (denoting zero refraction).
With d = 0, sampling will be uniform;
with d < 0, sample rate will decrease with distance from the origin. The sampling pattern will by optimized by evaluating and maximizing image-quality performance curves as a function of d.
Approaches to intermediate evaluations of image quality As explained in the previous sections, each of the image processing algorithms and data-acquisition protocols will be evaluated to optimize image quality and narrow the parameter space. In this section, we explain how image quality will be assessed in each case. Because of the large number of parameters and algorithms to be studied, it will not be practical to use full-scale radiologist reader studies as a day-to-day guidepost for the research. Therefore, in these __--a-f .--.

preliminary evaluations we will use: 1) regular feedback from one expert reader, Dr. Pisano, at the University of North Carolina; 2) standard quantitative image-quality measures, including numerical-observer detectability measures; and 4) periodic receiver-operating characteristic (ROC) reader studies involving the IIT team and their laboratory staff members.
Quantitative measures Standard quantitative measures of image quality will be used in preliminary evaluations. These will include bias-variance curves, contrast-to-noise ratios, and numerical observer performance.
These statistics all capture noise performance of the imaging system, and may be computed by three approaches: 1) by using many noise realizations from real data; 2) by using many noise realizations generated by computer using a theoretical noise model; or 3) by computing the statistics analytically using a theoretical noise model. Method 1 is, of course, the most faithful to the truth; however, the statistical estimates may have high variance due to the difficulty of obtaining large quantities of noise realizations. Method 2 overcomes the problem of insufficient noise realizations, provided that the noise model is carefully validated. With a well-validated noise model, Method 3 is ideal, because it not only produces the desired image-quality measures, but also provides analytical expressions for these measures in terms of important imaging parameters. Thus, Method 3 provides the greatest insights. We will aim to use Method 3, with experimental validation of noise models using actual noise data obtained at BNL using a step-wedge absorber. As explained in the Preliminary Results section, we have already developed and validated a theoretical noise model for DEI [2]. In this project, we will repeat the same process for MIR, using similar tools.
We will also develop numerical observers (statistical classifiers that mimic human object-detection performance) as a substitute for human readers in initial studies. Numerical observers, such as the channelized Hotelling observer (CHO), which have been pioneered by H.H. Barrett and others (e.g., [42]), have been widely and successfully used in various evaluation and optimization studies. Our group has significant experience with these observers, having developed them for modeling of perfusion-defect detection in cardiac imaging [43],[44]. We will develop numerical observers for detection of the simulated fibril and microcalcification features in the phantom from Aim 1, and use these observers as measures of object detectability in intermediate image-quality evaluations.
Reader studies using physicists and engineers As a periodic guidepost in the research, we will use members of the scientific team as readers in studies to measure visual detectability of image features, in the presence of noise, in a two-alternative, forced-choice detection task. A
simple graphical user interface (GUI) will be developed to allow the readers to view and score the images. These studies will be based on regions of interest (ROIs) in images of the phantom developed in Aim 1, so non-radiologists will be satisfactory readers for these studies. In the studies, the readers will be blinded to the presence or absence of each image feature (e.g., a phantom fibril or microcalcification of some particular size). Images will be presented on a softcopy high-luminance display (Clinton Electronics 21 inch, 8 megapixel monitor with Dome PCI Controller Card, Loves Park, IL) in a randomized and counterbalanced order to avoid biases due to order effects. Readers will report their confidence that an object is present in the ROI on a scale of 1-7 (1=object definitely absent; 2=object almost certainly absent; 3=object probably absent; 4=object possibly present;
5=object probably present; 6=object almost certainly present; 7=object definitely present). Reader scores will be analyzed by the software suite [40]
created by Prof. Charles Metz and colleagues at the University of Chicago (UofC) which estimates receiver operating characteristic (ROC) curves, area under the ROC curve (AUC), and estimates of significance of pairwise comparisons between treatments (image-processing algorithms, data-acquisition parameter choices, etc.) [41]. The UofC software uses jackknifing and analysis of variance (ANOVA) to account for reader and case variation in its comparisons. Bonferroni correction for multiple comparisons will be used when determining rejection of null hypothesis of no difference between treatments. At the outset of these evaluations, initial pilot studies will be performed to determine typical values of the binormal parameters that may be encountered in the study, so as to determine the number of readers and images required to achieve statistical significance of our comparisons. Since reader time will be contributed by members of our research group, we expect to have no trouble achieving statistical significance in comparisons where the magnitude of difference is large enough to be of practical interest.
D.4 Aim 3: Final reader study to compare MIR against conventional radiography The principal goals of the formal expert reader study will be to determine whether MIR images allow readers to see more useful diagnostic information than that provided by standard digital radiographs, and which of the three MIR
images is most useful (by the same criterion). We will also conduct an exploratory ROC study of discrimination between malignant and benign ROIs.

D.4.1 MIR images and radiographs A total of 24 breast specimens will be imaged. Twelve of the specimens will contain masses, and twelve will contain clusters of calcifications. Six of each type of lesion will be malignant. Therefore, images will be obtained of 12 cancers and 12 non-cancers, six of each type (that is, six cancerous masses, six noncancerous masses, six cancerous clusters of calcifications and six non-cancerous clusters of calcifications).
Planar images of these specimens will be obtained at the two best MIR
software/hardware configurations at Brookhaven National Laboratory as determined by the preliminary non-expert reader analysis performed at IIT.
Images will be emulated at BNL by adjusting x-ray fluence at the detector to match that expected from the designs generated in the optimization research. Images of the same specimens with the same orientations will be obtained at UNC using a General Electric Senographe 2000D digital mammography system.
D.4.2 Exploratory ROC analysis During the reader study, 10 expert radiologists will evaluate the images of all specimens using MIR absorption, refraction and scatter modes. Thus, the readers will each evaluate a total of 24x2x3 images for a total of 144 images.
The first task for the readers will be for them to view all of the MIR images in counterbalanced sequence and to assess the probability of malignancy on a 7-point scale (definitely malignant, almost certainly malignant, probably malignant, possibly malignant, probably not malignant, almost certainly not malignant, and definitely not malignant) and to give their probability of malignancy on a percentage (100 point) scale.
Once this is complete, we will perform an exploratory ROC analysis, including AUC. This part of the study may not result in a powerful test;
however, we believe the data are worth collecting before the readers are provided with the pathology report for the ROIs, which will occur in the next step of the study.
D.4.3 Reader preference study After the readers' first task is completed, the readers will be shown the digital radiograph of the same specimen with several regions of interest demonstrated by means of a transparent overlay with circles drawn on it. These areas will be identified by Dr. Etta Pisano, an expert breast imager, on the sets of specimen images created for this study and will be known as the regions of interest (ROI). Based on our prior experience, we expect 1 to 3 regions of interest per specimen or 24-72 regions of interest for al124 specimens. Drs. Pisano and Livasy, an expert breast pathologist, will together evaluate the regions of interest with reference to the histopathologic features that actually are visible in histologic whole mount slides of those areas.
From our prior experience, we expect all areas identified by Dr. Pisano as showing significant image detail have proven to reflect actual anatomic findings of interest [12]. Thus, there will be a total of 144-432 regions of interest for the readers in the 144 images to evaluate in this reader study.
The 10 radiologist readers who are expert in breast imaging will evaluate the visibility of the lesion details in the 144-432 different regions of interest in the 144 images of each specimen.
The radiologists will be given a detailed pathology report on each region of interest, which we expect to involve approximately 2-4 square centimeters each. Every image of every ROI will be ranked versus its visibility on the standard digital radiograph on a 7-point scale.
For example, if a spiculation that proved to be an extension of an invasive ductal carcinoma into surrounding tissue is a region of interest for specimen 1, a cancerous mass, the radiologists will be shown that portion of the image on the standard radiograph of that specimen and then asked to rate the MIR images for the visibility of that particular feature and their ability to see the feature on the experimental images versus the digital radiograph of the same image. In this manner, the MIR images that are best able to help the radiologists in evaluating the lesion's most important features can be identified.
This same process will take place for each ROI. Thus, since we expect 1-3 ROIs per each of the 24 specimens, the radiologists will rate the visibility of a total of 144-432 ROIs relative to the standard digital radiograph of the same specimens.
The order of image presentation to the radiologists will be counterbalanced to compensate for the effects of fatigue and experience. In addition, the radiologists will be assisted by a graduate research assistant who will aid them in completing the research forms and sorting through the images and ordering them appropriately. All reader studies will take place in a special room designed with very low ambient light, and very low probability of interruption, used specifically for reader studies at UNC. The images will be displayed on a softcopy high-luminance display (Clinton Electronics 21 inch, 8 megapixel monitor with Dome PCI Controller Card, Loves Park, IL). The radiologists will use a magnifying glass, as needed, as is standard clinical practice. Finally, the readers will _..- ;~--7 ,_,.

be required to take 10 minute breaks every 45 minutes. Prior to participation in this reader study, the 10 readers will be trained in the interpretation of MIR images by viewing the cases that were used to improve the system parameters during the first three years of the project.
Once the reader-data collection is complete, we will conduct statistical analyses. For each ROI, the ratings will be organized into a 10x2x3 matrix, corresponding to 10 readers x 2 imaging configurations x 3 MIR images. For the 24-72 ROIs (24 specimens, 1-3 ROIs per specimen), we will obtain 24-72 such matrices. The between-subject factors are the types of the specimen (i.e., malignant/benign and mass/calcification cluster). Next, repeated-measures ANOVA will be applied to analyze these data, and multiple tests with Bonferroni correction will be conducted to test the hypothesis that the images from MIR are preferred to the standard radiographs under each of the following combined conditions: 2 configurations x 3 MIR images x 4 specimen types.
In the following analysis, we evaluate the expected power of the test in the most conservative scenario. The worst-case scenario for Type I error is when each specimen yields three ROIs;
the worst case for power occurs when each specimen yields only one ROI. In this combined worst-case scenario, controlling overall Type I error to <0.05, the expected power of the test is as follows for some representative possible outcomes. If the mean rating is 0.5 times its standard deviation, then the power will be >0.75. If the mean rating is 0.6 times its standard deviation, then the power will be >0.93.
Once the best MIR images for improved lesion conspicuity are identified, later reader studies will be performed (under later grant funding), to determine how the readers rate combinations of the MIR images compared to the baseline condition, i.e. the digital mammogram, and how they utilize three-dimensionally acquired and displayed MIR data (MICT).
D.5 Aim 4: Compile specification for a clinical MIR breast imager The hardware design effort will lead to Final Design Reports (FDR) for each of the four hardware configurations. The FDR for each hardware configuration will then be augmented with the following corresponding information: 1) the recommended data-acquisition procedure; 2) the recommended image-processing procedures; 3) the quantitative image-quality results; and 4) the statistical analyses of the final radiologist reader studies for two selected design configurations. These FDRs will form the basis for a proposal to construct a prototype imaging system in future work.

~. a~ ~

G. LITERATURE CITED

[1] D. Chapman, W. Thomlinson, R.E. Johnston, D. Washburn, E. Pisano, N. Gmur, Z. Zhong, R. Menk, F. Arfelli, D. Sayers, "Diffraction Enhanced X-ray Imaging", Phys. Med. Biol., vol. 42, pp. 2015-2025, 1997.
[2] M. N. Wernick, O. Wirjadi, D. Chapman, O. Oltulu, Z. Zhong, and Y. Yang, "Preliminary investigation of a multiple-image radiography method," IEEE/NIHIntI. Symp. Biomed. Imaging, pp.
129-132, 2002.
[3] O. Oltulu, Z. Zhong, M. Hasnah, M. N. Wernick, and D. Chapman, "Extraction of extinction, refraction and absorption properties in diffraction enhanced imaging," J Phys. D: Appl.
Phys., vol. 36, pp. 2152-2156, 2003.
[4] M. N. Wernick, O. Wirjadi, D. Chapman, Z. Zhong, N. P.Galatsanos, Y. Yang, J. G. Brankov, O. Oltulu, M.
A.Anastasio, and C. Muehleman, "Multiple-image radiography," Phys. Med. Biol., vol. 48, pp. 3875-3895, 2003.
[5] C. Muehleman, J.G. Brankov, J. Li, Z. Zhong, K.E. Kuettner, and M.N.
Wernick, "Multiple-image radiography for soft tissue," Proceeding of the Orthopedic Research Society, No. 104, 2004.

[6] N. Yagi, Y. Suzuki, K. Umetaani, Y. Kohmura, Y. Yamasaki, "Refraction-enhanced x-ray imaging of mouse lung using synchrotron radiation source," Med. Phys. 26(10), 2190-2193 (1999).

[7] F. Arfelli, V. Bonvicini, A. Bravin, G. Cantatore, E. Castelli, L.D.
Palma, M.D. Michiel, M. Fabrizioli, R.
Longo, R.H. Menk, A. Olivio, S. Pani, D. Pontoni, P. Poropat, M. Prest, A.
Rashevsky, M. Ratti, L. Rigon, G.
Tromba, A. Vacchi, E. Vallazza, F. Zanconati, "Mammography with Synchrotron Radiation: Phase-Detection Techniques," Radiology 215, 286-293 (2000).

[8] S. Fiedler, A. Bravin, J. Keyrilainen, M. Fernandez, P. Suortti, W.
Thomlinson, M. Tenhunen, P. Virkkunen, and M-L Karjalainen-Lindsberg, "Imaging lobular breast carcinoma: comparison of synchrotron radiation DEI-CT technique with clinical CT, mammography and histology," Phys. Med. Biol., 49:175-188, 2004.

[9] E. Pagot, P. Cloetens, S. Fiedler, A. Bravin, P. Coan, J. Baruchel, J.
Hartwig, and W. Thomlinson, "A method to extract quantitative information in analyzer-based x-ray phase contrast imaging," Applied Physics Letters, 82:3421-3423, 2002.

[10] J. Mollenhauer, A.E. Aurich, Z. Zhong, C. Muehleman, A.A. Cole, M.
Hasnah, O. Oltulu, K.E. Kuettner, A.
Margulis, L.D. Chapman, "Diffraction-enhanced X-ray imaging of articular cartilage," Osteoarthritis and Cartilage / OARS, vol. 10(3), pp. 163-71, 2002.

[11] M. Hasnah, Z. Zhong, O. Oltulu, E. Pisano, R.E. Johnston, D. Sayers, W.
Thomlinson and D. Chapman "Diffraction enhanced imaging contrast mechanisms in breast cancer specimens,"
Medical Physics, vol. 29, no.
10, October 2002.

[12] E.D. Pisano, R.E. Johnston, D. Chapman, J. Geradts, M.V. Iacocca, C.A.
Livasy, D.B. Washburn, D.E. Sayers, Z. Zhong, M.Z. Kiss, and W. Thomlinson, "Diffraction enhanced irnaging of human breast cancer specimens:
improved conspicuity of lesion detail with histologic correlation," Radiology 2000, 214, 895-901.

[13] "DEI Compression Study," U.S. Army Medical Research and Materiel Command grant DAMD17-02-0523;
Principal Investigator: C. A. Parham.

[14] R.E. Hendrick, R.A. Smith, J.H. Rutledge, C.R. Smart, "Benefit of screening mammography in women ages 40-49: a new meta-analysis of randomized controlled trials," Journal of the National Cancer Institute Monograph, vol. 22, pp. 87-92, 1997.

[15] T.A. Graves and K.I. Bland, "Surgery for early and minimally invasive breast cancer," Current Opinions in Oncology, vol. 8(6), pp. 468-477, 1996.

[16] M.B. Laya, E.B. Larson, S.H. Taplin, E. White, "Effect of estrogen replacement therapy on the specificity and sensitivity of screening mammography," JNCI, vol. 88(10), pp. 643-649, 1996.

[17] L. H. Baker, "Breast cancer detection demonstration project: five-year summary report," CA, vol. 32(4), pp.
194-225, 1982.

[18] Sickles E. Periodic mammographic follow-up of probably benign lesions:
Results in 3,184 consecutive cases,"
Radiology, vol. 179(2), pp. 463-468, 1991.

[19] D.M. Ikeda, R.L. Birdwell, K.F. O'Shaughnessy, R.J. Brenner, E.A.
Sickles, "Analysis of 172 subtle findings on prior normal mammograms in women with breast cancer detected at follow-up screening," Radiology, vol.
226(2), pp. 494-503, 2003.

[20] E.D. Pisano and M. Yaffe, "Digital mammography: State of the art,"
Radiology, in press.
[21 ] T.S. Mehta, "Current uses of ultrasound in the evaluation of the breast," Radiol Clin North Am., vol. 41(4), pp.
841-856, 2003 [22] E.A. Morris, "Screening for breast cancer with MRI," Semin Ultrasound CT
MR, vol. 24(1), pp. 45-54. 2003.
[23] U. Bonse and M. Hart, Appl. Phys. Lett. vol. 7, pp. 238, 1965.
[24] M. Hasnah, Z. Zhong, O. Oltulu, E. Pisano, R.E. Johnston, D. Sayers, W.
Thomlinson and D. Chapman "Diffraction enhanced imaging contrast mechanisms in breast cancer specimens,"
Med Phys, vol. 29, no. 10, October 2002.
[25] M. Yaffe, personal communication. According to Dr. Yaffe, in his experience based on thousands of clinical scans, 6-sec scans have not resulted in significant motion artifacts.
[26] J. G. Brankov, A. Saiz-Herranz, and M. N. Wernick, "Noise analysis for diffraction enhanced imaging," Proc IEEE/NIHIntl Symp Biomed Imaging, 2004. (full journal paper also submitted to Phys. Med. Biol.) [27] A. Saiz-Herranz, "Noise analysis of diffraction-enhanced imaging," M.S.
Thesis, Illinois Institute of Technology, 2003.
[28] G. J. McLaughlin and D. Peel, Finite Mixture Models, Wiley-Interscience, NY, 2000.
[29] H. Stark and Y. Yang, Vector Space Projections: A Numerical Approach to Signal and Image Processing, Neural Nets, and Optics, John Wiley & Sons, Inc., New York, 1998.
[30] G. Khelashvili, I. Ivanov, T.I. Morrison, G. Bunker, D. Chapman, "A novel method of bending crystals to log spiral shape," Review of Scientific Instruments vol. 73, pp. 1534-6, 2002.
[31] J A Halbleib, R P Kensek, T A Mehlhorn, G D Valdez, S M Seltzer and M J
Berger, "ITS Version 3.0: the integrated Tiger series of coupled electron/photon Monte Carlo transport codes," Report SAND 91-1634 (Sandia National Laboratory), 1992.
[32] J A Halbleib, R P Kensek, T A Mehlhorn, G D Valdez, S M Seltzer and M J
Berger, "ITS: the integrated TIGER series of electron/photon transport codes-version 3.0," IEEE Trans.
Nucl. Sci., vol. 39, pp. 1025-1030, [33] A. P. Dempster, N. M. Laird, and D. B. Rubin, "Maximum likelihood from incomplete data via the EM
algorithm," J. Roy. Statist. Sect., vol. 39, pp. 1-38, 1977.
[34] D.S. Lalush and M.N. Wernick, "Iterative Image Reconstruction," in Emission Tomography: The Fundamentals of PET and SPECT, M.N. Wernick and J.N. Aarsvold, eds., San Diego: Academic Press, 2003, in press.
[35] N. P. Galatsanos, M. N. Wernick, and A. K. Katsaggelos, "Multichannel Image Recovery," in Handbook of Image and Video Processing, A. Bovik, Ed., San Diego: Academic Press, pp. 155-168, 2000.
[36] C.-M. Kao, J. T. Yap, J. Mukherjee, and M. N. Wernick, "Image reconstruction for dynamic PET based on low-order approximation and restoration of the sinogram," IEEE Trans. Med.
Imaging, vol. 16, pp. 738-749, 1997.
[37] M. N. Wernick, E. J. Infusino, and M. Milosevic, "Fast spatio-temporal image reconstruction for dynamic PET," IEEE Transactions on Medical Imaging, vol. 18, pp. 185-195, March 1999.
[38] V. M. Narayanan, M. A. King, E. Soares, C. Byrne, H. Pretorius and M. N.
Wernick, "Application of the Karhunen-Loeve transform to 4D reconstruction of gated cardiac SPECT images,"
IEEE Trans. Nucl. Sci., vol.
46, pp. 1001-1008, 1999.

C) [39] V. M. Narayanan, M. A. King, M. N. Wernick, et al, "Improved image quality and computation reduction in 4D
reconstruction of cardiac gated SPECT images," IEEE Trans. Med. Imag., vol.
19, 2000.
[40] ROC software, Kurt Rossmann Laboratories for Radiologic Image Research, http://www-radiology. uchicago.edu/krl/roc_soft.htm.
[41] D.D. Dorfman, K.S. Berbaum, C.E. Metz, "ROC rating analysis:
generalization to the population of readers and cases with the jackknife method," Invest Radiol, vol. 27, pp. 723, 1992.
[42] K. J. Myers and H. H. Barrett, J. Opt. Soc. Am. A, vol. 4, no. 12, 1987.
[43] M.V. Narayanan, H.C. Gifford, M.A. King, P.H. Pretorius, T.H. Farncombe, P. Bruyant, and M.N. Wernick, "Optimization of iterative reconstructions of Tc-99m cardiac SPECT studies using numerical observers," IEEE
Transactions on Nuclear Science, vol. 49, pp. 2355-2360, 2002.
[44] J. G. Brankov, I. El Naqa, Y. Yang, M. N. Wernick, "Learning a nonlinear channelized observer for image quality assessment," Proc IEEE Med Imag Conf, 2003.

2 1 __.-

Claims