Embed Size (px)
Citation preview
Tissue Equivalent Phantom Design for Characterization of a Coherent Scatter X-ray
Imaging System
by
Kathryn Elizabeth Albanese
Medical Physics Graduate Program Duke University
Date: _____________________
Approved:
___________________________ Anuj Kapadia, Supervisor
___________________________
Joel Greenberg
___________________________ Joseph Lo
___________________________
Robert Reiman
Thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in the Medical Physics Graduate
Program in the Graduate School of Duke University
2016
Tissue Equivalent Phantom Design for Characterization of a Coherent Scatter X-ray
Imaging System
by
Kathryn Elizabeth Albanese
Medical Physics Graduate Program Duke University
Date: _____________________
Approved:
___________________________ Anuj Kapadia, Supervisor
___________________________
Joel Greenberg
___________________________ Joseph Lo
___________________________
Robert Reiman
An abstract of a thesis submitted in partial fulfillment of the requirements for the degree
of Master of Science in the Medical Physics Graduate Program in the Graduate School of Duke University
2016
Copyright by Kathryn E. Albanese
2016
iv
Abstract Scatter in medical imaging is typically cast off as image-related noise that
detracts from meaningful diagnosis. It is therefore typically rejected or removed from
medical images. However, it has been found that every material, including cancerous
tissue, has a unique X-ray coherent scatter signature that can be used to identify that
material or tissue. Such scatter-based tissue-identification provides the advantage of
material-specific identification in 3D over conventional anatomical imaging through X-
ray radiography. A coded aperture X-ray coherent scatter spectral imaging system has
been developed in our group to classify different tissue types based on their unique
scatter signatures. Previous experiments using our prototype have demonstrated that
the depth-resolved coherent scatter spectral imaging system (CACSSI) can discriminate
healthy and cancerous tissue present in the path of a non-destructive X-ray beam. A key
to the successful optimization of CACSSI as a clinical imaging method is to obtain
anatomically accurate phantoms of the human body. This thesis describes the
development and fabrication of 3D printed anatomical scatter phantoms of the breast
and lung.
The purpose of this work is to accurately model different breast geometries using
a tissue equivalent phantom, and to classify these tissues in a coherent X-ray scatter
imaging system. Tissue-equivalent anatomical phantoms were designed to assess the
v
capability of the CACSSI system to classify different types of breast tissue (adipose,
fibroglandular, malignant). These phantoms were 3D printed based on DICOM data
obtained from CT scans of lungs and prone breasts. The phantoms were tested through
comparison of measured scatter signatures with those of adipose and fibroglandular
tissue from literature. Tumors in the phantom were modeled using a variety of
biological tissue including actual surgically excised benign and malignant tissue
specimens. Lung based phantoms have also been printed for future testing. Our imaging
system has been able to define the location and composition of the various materials in
the phantom. These phantoms were used to characterize the CACSSI system in terms of
beam width and imaging technique. The result of this work showed accurate modeling
and characterization of the phantoms through comparison of the tissue-equivalent form
factors to those from literature. The physical construction of the phantoms, based on
actual patient anatomy, was validated using mammography and computed tomography
to visually compare the clinical images to those of actual patient anatomy.
vi
Dedication To my family for always supporting me in the pursuit of my dreams, and to my
friends for still liking me from 522 miles away. All the love.
vii
Contents
Abstract .......................................................................................................................................... iv
Dedication ..................................................................................................................................... vi
List of Tables ................................................................................................................................. ix
List of Figures ............................................................................................................................... xi
Abbreviations ............................................................................................................................. xiv
Acknowledgements .................................................................................................................... xv
1. Introduction ............................................................................................................................... 1
1.1 Breast Cancer Pathology .................................................................................................. 2
1.1.1 Value of Coherent Scatter Imaging in Breast Cancer Diagnosis .......................... 4
1.2 Coherent Scatter ................................................................................................................ 5
1.2.1 Coded Aperture Coherent Scatter Spectral Imaging ............................................. 7
1.2.2 Motivation for Using CACCSI ................................................................................... 8
1.3 Anthropomorphic Phantoms .......................................................................................... 9
2. Phantom Development .......................................................................................................... 11
2.1 Calibration Phantoms .................................................................................................... 12
2.2 Scatter Equivalent Materials as Tissue Surrogates .................................................... 12
2.1.1 Previous Work ........................................................................................................... 21
2.3 Early Phantom Models .................................................................................................. 22
2.4 Current Phantom Models .............................................................................................. 23
2.4.1 Prone Breast Geometry ............................................................................................. 24
viii
2.4.2 The 3D Printing Process ........................................................................................... 25
2.4.3 Segmentation of DICOM Data ................................................................................ 28
3. Phantom Validation and Verification .................................................................................. 31
3.1 Validation of Anatomy .................................................................................................. 31
3.2 Form Factors .................................................................................................................... 34
3.3 Scanning Process ............................................................................................................ 34
3.3.1 Raw Data .................................................................................................................... 37
3.3.2 Reconstruction ........................................................................................................... 38
3.4 Verification Methodology ............................................................................................. 40
4. Results and Discussion ........................................................................................................... 43
4.1 Metrics of Comparison .................................................................................................. 43
5. Conclusions and Future Work .............................................................................................. 55
5.1 Pathology Protocol ......................................................................................................... 55
5.2 Printing Filament Experimentation ............................................................................. 56
5.3 Solid 3D Printed Phantom ............................................................................................. 57
References .................................................................................................................................... 58
ix
List of Tables Table 1: A description of materials used within the phantom and the purpose they serve......................................................................................................................................................... 13
Table 2: This table shows the correlation between the human tissue and the tissue surrogates used within the phantom. ...................................................................................... 15
Table 3: A table of correlation values between materials used in the phantom as well as the tissues they mimicked. ......................................................................................................... 18
Table 4: Correlations of different materials with the adipose form factor. Lard is the one chosen as the surrogate material. ............................................................................................. 19
Table 5: Correlation of different materials with the cancer form factor. Actual excised human tissue, shown below, was used as the surrogate for cancerous tissue in the majority of phantom work. ........................................................................................................ 19
Table 6: Correlation of different materials with the fibroglandular form factor. As normal tissue is a mixture of adipose and fibroglandular tissue, the healthy excised human tissue was considered a surrogate for fibroglandular tissue. ............................................... 20
Table 7: Correlation of different materials with the normal tissue form factor. Normal tissue is a mixture of adipose and fibroglandular tissue and the excised human tissues were used as normal tissue. ....................................................................................................... 20
Table 8: A table of the Hounsfield Numbers of different materials .................................... 33
Table 9: A table of materials scanned and metrics of comparison used in analysis. The materials were first scanned alone in the beam, and then surrounded by lard within the phantom. Several of the metrics of comparison were simply visual: number of major peaks, overall peak shape and peak shift. Peak shift indicates a shift of the material form factor toward the peak of lard when placed inside the phantom. The peak location along the x-axis is both a visual and quantitative value. The correlation of the material alone and within the phantom is a quantitative metric described in a later section. .................. 42
Table 10: Lard metric results ..................................................................................................... 45
Table 11: Beef metric results. ..................................................................................................... 46
x
Table 12: Water metric results. .................................................................................................. 49
Table 13: Tumor metric results. ................................................................................................. 50
Table 14: Normal metric results. ............................................................................................... 52
Table 15: This table shows the correlation between the form factor of a material with the form factor of the same material placed in the phantom with lard. .................................... 52
xi
List of Figures Figure 1: Examples of scatter signatures (form factors) as measured by Kidane, et al. [3] 7
Figure 2: The coded aperture mask where each element is preferentially attenuating. ..... 7
Figure 3: Powdered metals were inserted into small Nalgene containers as calibration phantoms. ..................................................................................................................................... 12
Figure 4: A comparison of form factors from human tissues (dashed) and their phantom surrogates (solid). Of note is the similarity of peak location and spectral shape of the human materials and their surrogates. [34] ............................................................................ 14
Figure 5: A comparison of the scatter signatures from the 3D printing filament (blue), lard (green) and reconstructed lard (thin blue). The correlation between lard and plastic form factors is 0 .953, meaning the plastic would be classified as lard within the CACSSI reconstruction. ............................................................................................................................. 17
Figure 6: This version of the phantom included “shelves” for the convenient placement of excised human tissue samples or calibration phantoms. ................................................. 22
Figure 7: A SketchUp rendering of a DICOM modeled prone breast phantom ................ 23
Figure 8: The 3D printed model of the rendering in figure 1. .............................................. 24
Figure 9: A mammogram of the phantom in the horizontal plane shows the hexagonal infill structure of the phantom shell, which presents itself as vertical lines in a coronal or sagittal view. ................................................................................................................................ 24
Figure 10: An image of the ITK-Snap interface while segmenting the lungs using the classification trainer. ................................................................................................................... 30
Figure 11: The printed phantom from the segmentation shown in figure 10. ................... 30
Figure 12: The setup used to take a mammogram of the phantom. .................................... 32
Figure 13: The resulting mammographic image. The “streaks” in the image are due to the plastic infill within the phantom shell. .............................................................................. 32
Figure 14: A CT scan of the phantom filled with lard. The phantom on the right also contains a Nalgene filled with graphite. .................................................................................. 33
xii
Figure 15: The scanning geometry. A is the x-ray tube, B is a collimator, C is the object stage, D is the coded aperture and E is the detector. The samples are scanned using 125 kVp and 100 mAs with a ~1mm collimated pencil beam.. [34,35] ....................................... 36
Figure 16: An example of the bell shaped curve generated in the raw data with y-axis representing Energy [keV] and x-axis representing pixel number. ..................................... 37
Figure 17: This is the first plot shown in the reconstruction code. This example was taken from a pencil beam scan of lard alone. The user clicks the spot on the z-axis where the object was known to be placed (which usually corresponds to the brightest white pixel in this image). That gives the reconstructed form factors through that pixel. ......... 39
Figure 18: This figure shows the distribution of tissue types along the beam axis from the same scan as figure 15 above. The color map indicates that all of the q values detected by the system matched that of lard. ............................................................................................... 39
Figure 19: Raw data from scans taken of (a) background signal, (b) the vacuum bags, (c) chicken breast and (d) chicken breast inside the vacuum bags. The decrease in counts from (a) to (b) shows self attenuation of the bags. ................................................................. 41
Figure 20: This figure shows the subtraction of the chicken alone data (18.c) and the chicken in the vacuum bag (18.d). Rather than indicating a strong scatter pattern, this subtracted image shows random noise that indicates the bags do not have a strong signal in the scatter imaging system. ....................................................................................... 41
Figure 21: Reconstructed form factor of lard within the phantom geometry (blue) and lard alone (solid green). The correlation is .983. ..................................................................... 44
Figure 22: Reconstructed form factors of dried beef alone (left) and dried beef inside the breast phantom filled with lard (right). Of note is the dual peaks on the right which indicate the presence of both lard and beef. ........................................................................... 45
Figure 23: Extracted reconstructed form factors of beef alone (blue) and beef inside lard in the phantom (red) plotted together for comparison. Their correlation is .911. ............. 46
Figure 24: As mentioned in section 1.2.1, it was determined that the order of materials placed within the beam does not affect their ability to be detected. The image on the left is the form factor of beef followed by lard, and the image on the right is the form factor of lard followed by beef within the beam. .............................................................................. 47
xiii
Figure 25: Extracted reconstructed form factors of beef followed by lard (blue) and lard followed by beef in the phantom (red) plotted together for comparison. Their correlation is .925 and shows that the signal differs mostly in amplitude rather than shape or peak location. This demonstrates that the ordering of objects within the beam does not affect their ability to identified. ........................................................................................................... 47
Figure 26: Reconstructed form factors of water alone (left) and water inside the breast phantom filled with lard (right). Water was used as a calibration material. Of note is the shift on the right of the reconstructed form factor (blue) toward the lard form factor, which indicate the presence of both lard and water in the beam. ....................................... 48
Figure 27: Extracted reconstructed form factors of water alone (blue) and water inside lard in the phantom (red) plotted together for comparison. Their correlation is .891. ..... 48
Figure 28: Reconstructed form factors of tumor alone (left) and tumor in lard inside the phantom (right). Of note is the change in reconstructed form factor shape between the two plots, the second reconstruction more closely mimicking the shape of the lard form factor. ............................................................................................................................................ 49
Figure 29: Extracted reconstructed form factors of tumor alone (blue) and tumor inside lard in the phantom (red) plotted together for comparison. Their correlation is .985. ..... 50
Figure 30: Reconstructed form factors of normal tissue alone (left) and normal tissue in lard inside the phantom (right). Of note is the change in reconstructed form factor shape between the two plots, the second reconstruction more closely mimicking the shape of the lard form factor. .................................................................................................................... 51
Figure 31: Extracted reconstructed form factors of normal tissue alone (red) and normal tissue inside lard in the phantom (blue) plotted together for comparison. Their correlation is .885. ....................................................................................................................... 51
Figure 32: (a) A tumor classification map rendered by CACSSI. (b) An interpolated image to register with the histological report in (c). [35] This is an example of how the form factor from each voxel is matched to a spectrum from the form factor library through correlation and classified as a material. ................................................................... 54
xiv
Abbreviations ABS- Acylonitrile Butadiene Styrene
BCS- Breast Conserving Surgery
CACSSI- Coded Aperture Coherent Scatter Spectral Imaging
CAD- computer aided design
CAM- computer aided manufacturing
CNC- computerized numeric control
CSCT- Coherent Scatter Computed Tomography
DICOM- Digital Imaging and Communications in Medicine
EDXRD- Energy Dispersive X-ray diffraction
ECM- Extra Cellular Matrix
FDM- Fused Deposition Modeling
H&E- Hematoxylin and eosin stain
HER2- Human Epidermal Growth Factor Receptor 2
IRB- institutional review board
MAP- maximum a posteriori
PLA- Poly Lactic Acid
SBMT- skull-base to mid-thigh
TV- total variation
XCAT- eXtended CArdia-Torso phantom
xv
Acknowledgements
First and foremost, I would like to thank my advisor, Anuj Kapadia, for his
unwavering support and encouragement throughout my graduate career. His
mentorship has made me not only a better researcher, but a stronger person. He brought
out the best in me by seeing me through my worst, and for that I will be forever grateful.
Joel Greenberg and Manu Lakshmanan have been invaluable to the research I
have done. The hours they spent educating me about scatter imaging helped me truly
appreciate the value of the work being done. Their willingness to help guide my writing
and conference submissions was greatly appreciated.
I am so lucky to have had Robert Morris as my partner in crime throughout my
experience at Duke. Robert made everything more fun- even Raster Scan Day. There is
nobody I would have rather had by my side during lab meetings, conferences and in D2.
I’m so happy I could finally print your lungs for you. They’re truly magnificent.
Through our research partnership, I have made a lifelong friend.
Research is a collaborative effort, and there are so many others I need to thank
for helping me in this thesis work. James Spencer, although I only got to work with him
for a few months in the lab, made D2 a better place. His eagerness to help out and get
involved in all aspects of research made writing abstracts and this thesis easier and more
enjoyable. Devin Miles helped to design our first phantoms and taught me how to 3D
xvi
print and use segmentation software, which was the foundation of my research. Chip
Bobbert and the rest of the Innovation Studio put up with several hundred hours of my
prints and taught me the fundamentals of how 3D printing can best be utilized. Martin
Tornai provided the breast CT data that was used to model the phantoms, and was
always willing to assist in any way he could. Paul Yoon provided code to convert file
types for segmentation. Ehsan Samei supplied the lab space where D2 was housed, and
the Clinical Imaging Physics Group, namely Jeff Nelson and James Winslow, facilitated
the clinical scans of the phantoms. Paul Segars supplied the XCAT data from which I
derived lung phantoms. Nicole Ball gave me tools to perfect my 3D printed items. Scott
Wolter assisted in the X-ray diffraction work and always welcomed us into his lab at
Elon. Mridu Nanda spent her school breaks in our lab working on the automation of the
various moving parts.
Working in the RAILabs fostered an environment for collaboration and was a
valued resource for creativity in research. Thanks to the other members of my
committee, Joseph Lo and Robert Reiman, for their input and guidance in RAILabs
Friday Forums and classes.
1
1. Introduction In the United States, about one in eight women will develop invasive breast
cancer and about one in thirty-six will die of it [1], making research and treatment of this
disease a very important task. Clinical practice in breast imaging uses mammography
for breast cancer screening and diagnosis and might also include CT, MRI or Ultrasound
in the diagnostic workup. Detection and diagnosis of breast cancer through imaging is a
difficult task in any modality due to the subtle attenuation differences between normal
and cancerous tissue as well as the small structures, such as microcalcifications, that are
often indicative of breast cancer. It has been found that coherent scatter imaging can
reveal properties about the molecular structure of a scanned sample rather than the
attenuation properties of that sample. Thus, a coherent scatter imaging system can be
used to identify different tissue types within a scanned object without taking into
account any attenuative differences of the tissues being scanned. Coherent scatter
imaging is used to measure a unique scatter signature for each scanned material and can
be compared in order to classify those materials as a certain type by correlating their
scatter signatures to a known library of scatter signatures. The aim of this work was to
create phantoms that model realistic anatomy and scatter properties of human tissue
and tumors. These phantoms are used to assess the capability of the scatter imaging
system to function in a clinical environment on realistic human geometries. Work done
by Kidane in measuring the scatter signatures of breast tissues using energy dispersive
2
X-ray diffraction (EDXRD) laid a foundation for identifying tissue-equivalent phantom
materials before using actual human tissue samples.
1.1 Breast Cancer Pathology
Breast imaging and cancer diagnosis is a clinically challenging task due to small
soft tissue contrast in the breast as well as the spiculated shape of breast tumors. This
irregular shape poses a difficulty in identifying the extent of breast cancer. Current
clinical practice for the treatment of breast cancer surgical intervention falls into one of
two types: breast conserving surgery (BCS) or mastectomy. Breast conserving surgery
only excises the malignant tumor and a small amount of normal surrounding tissue, and
is chosen by the patient in 59% of surgical cases [9] because survival rates rival that of
mastectomy but have favorable cosmetic outcomes. The margin of normal tissue
surrounding the tumor is necessary to determine whether the extent of the cancer was
removed from the body. These margins can be visually determined by the physician or
analyzed through other methods, such as the gold standard: histology analysis by a
pathologist. This type of analysis often takes days and often results in patient recall due
to findings of positive tumor margins. Additionally, many of the currently available
intra-operative margin detection techniques have sensitivities as low as 38% [10] and
increase the rate of patient recall unnecessarily. Repeat surgeries are a burden
physically, emotionally and financially for the patient and should thus be avoided. One
of the main objectives in assessing the compatibility of breast cancer detection in the
3
Coded Aperture Coherent Scatter Imaging (CACSSI) system was to determine whether
the system could be used in intra-operative margin detection for breast conserving
surgery, as investigated by Lakshmanan et al. [8]. There also exists the possibility of use
in-vivo to act as a complement to conventional radiography and mammography for
cancer screening and diagnosis and virtual biopsy.
The use of small specimens as well as low energy x-rays allows for multiple
scatter to be treated as low frequency background [38]. According to Batchelar et al,
multiple scatter can be ignored for a specimen less than three half value layers in
diameter [11]. This amounts to approximately 10cm for diagnostic X-ray energies in
human tissue. One challenge of this work was to utilize phantoms that more accurately
resembled human proportions to investigate the possibility of using CACSSI for larger
samples, and eventually in-vivo. Thus, the phantoms were significantly larger in depth
than 10cm. This caused an increased background signal in the image, but no noticeable
trouble in image interpretation as of yet. Multiple scatter also causes less distortion for
pencil beam geometry than for fan beam, the former being used in this system.
Cancer cell accumulation in a tissue is reflected in the tumor shape, resulting
from interaction of proliferating cells within their existing environment [12]. The PLOS
study defined four different types of tumor volumes- spheric, prolate, oblate and
ellipsoid by analyzing the tumor dimensions in three dimensions. This study found that
tumors such as those in triple negative breast cancer, exhibited spherical volumes with
4
well defined borders. Triple negative breast cancer means that none of the three
estrogen, progesterone and HER2 receptors that typically feed breast cancer
proliferation are present in the tumor. Therefore, these receptors cannot be targeted and
chemotherapy is often the most effective treatment option [13]. Otherwise, it was found
that the tumors with the most eccentric shapes demonstrated the worst prognosis. It was
also determined that the genes used in remodeling the extracellular matrix during
cancer proliferation are what likely causes heterogeneity in breast tumor structure [12].
The structure of the ECM contributes to the unique form factor of breast tumor from
other breast tissues, as coherent scatter signatures are dependent on differences in
intermolecular spacing. We now understand that the genes that build the ECM are
different from those within other breast tissues and allow us to distinguish these tissues
in our system.
1.1.1 Value of Coherent Scatter Imaging in Breast Cancer Diagnosis
Coherent scatter studies in breast cancer [3, 10, 14, 15, 16] have shown that the
differential coherent scatter cross sections of tissues inherent in the breast are unique
enough to identify those materials based on the angular distribution of coherent scatter
coming from them [10, 17, 18, 19, 20, 21, 22]. This is due in part to the disruption of the
extracellular matrix inherent in cancerous tissue. The distortion of the extracellular
matrix structure in cancer lends itself to identification by CACSSI [23 24] In our group,
5
CACSSI has been successfully used to discriminate adipose, fibroglandular, carcinoma
and normal breast tissue (which is a mix between adipose and fibroglandular tissue) [8].
The tissues of the breast have inherently low subject contrast, or ratio of
transmitted radiation intensities from different types of tissue, due to their small
differences in attenuation [25]. In addition, there is large variation in breast shape and
tissue composition among the population. CACSSI is also able to render a three-
dimensional reconstruction of the imaged specimen without the need for tomographic
rotation around the specimen. This offers a view deeper inside a specimen than other
experimental methods such as optical imaging [25, 26, 27, 28] and is helpful in
confirming the resection of multifocal disease. The work done with these phantoms
takes advantage of the depth resolution capabilities of CACSSI by experimenting with
human equivalent anatomical geometry.
1.2 Coherent Scatter
Coherent scatter, also known as Rayleigh or Small Angle Scatter, is one of the
methods through which X-rays interact with matter. It occurs when the energy of an
incident X-ray photon is small compared to the ionization energy of the target. Because
the energy of the incident photon is so low, negligible energy transfer to the target
medium occurs upon interaction. Rather, the incident photon experiences a change in
trajectory, i.e., “scatter”. Due to its relatively small cross section and similar appearance
to primary radiation, coherent scatter is often ignored in diagnostic radiology [2,3].
6
However, it has been found that the scatter properties of tissues are often unique and
can be used to classify them. In the breast, the attenuation properties of different breast
tissues are such that the contrast between tissues in mammographic images can make
breast lesions hard to detect. Rather than dismiss scatter as a phenomenon that degrades
the image quality for diagnostic breast imaging, Kidane et al [3] thought to use the
different scatter properties of each type of tissue as an identifier of sorts. The different
scatter signatures of the tissues are dependent on molecular composition and
intermolecular spacing because the x-ray wavelengths are consistent with the
dimensions of intermolecular spacing. [4,5].
The scatter signatures are measured via momentum transfer per nanometer (“q”)
in the scattered photon as given by Bragg’s Law:
𝑞 =12𝑑
=𝐸ℎ𝑐sin
𝜃2
where q is the momentum transfer, d is the effective lattice spacing of the
material, E is the X-ray energy, h is Planck’s constant, c is the speed of light in vacuum,
and 𝜃 is the angle of deflection of the X-ray [6]. These scatter signature functions are
known as form factors, proportional to the Fourier Transform of the electron charge
distribution of the sample [7]. When investigated in powder diffraction experiments,
diffracted X-rays appear as cones for each momentum transfer q on X-ray detectors [8].
This technique, as subsequently studied in medical and homeland security settings,
yields a higher SNR than using transmitted X-rays [7].
7
Figure 1: Examples of scatter signatures (form factors) as measured by Kidane, et al. [3]
1.2.1 Coded Aperture Coherent Scatter Spectral Imaging
The Coded Aperture Coherent Scatter Spectral Imaging system (CACSSI) uses a
collimated beam of X-rays, a sample stage, and moving translation stages that house the
detector mounts. What is unique about CACSSI is the use of a coded aperture mask to
facilitate a depth resolved image reconstruction.
Figure 2: The coded aperture mask where each element is preferentially attenuating.
8
The introduction of a coded aperture multiplexes the signal by “spatially
modulating scattered X-rays by a frequency spatially dependent on the origin of the
scatter” [8]. The aperture enables determination of position along the beam direction in
the z-axis. A multiplexed signal means that one detector pixel receives information from
several voxels within the scan sample. This sample is de-multiplexed using a forward
model in our reconstruction algorithm. Identifying the origin of the scatter enables
reconstruction without rotation of the sample. This contributes to the shorter scan time
than in a previous study investigating coherent scatter computed tomography (CSCT).
The CACSSI setup served as an experimental validation for a Monte Carlo study done
previously by Lakshmanan et al with a promising correlation coefficient between
measured and reconstructed form factors of .84 [8]. The same study also demonstrated
that the ordering of materials within the sample being scanned did not significantly
affect their ability to be detected [7]. Quantitative data shows this in section 4.1 of this
work.
1.2.2 Motivation for Using CACCSI
Lakshmanan et al [8] investigated the use of coherent scatter computed
tomography (CSCT) to image breast tissues. While the method was found effective, it
was determined to be prohibitive in the time required for the scan. They concluded that
CACSSI would reduce the scan time by an order of magnitude and would be of greater
9
advantage in clinical applications. The shorter scan time utilized by CACSSI also helps
minimize radiation dose as well as hardware fatigue (e.g., overheating of the X-ray
tube). Given the clear potential of CACSSI in clinical imaging, further investigation into
the optimization of the technique was pursued in this work – particularly with the
development of anatomically accurate phantoms for system optimization.
1.3 Anthropomorphic Phantoms
Human tissue-equivalent phantoms have been widely used in medicine for
dosimetry and quality control purposes. These phantoms typically mimic the
attenuation, density and effective atomic number characteristics of human tissue [41, 42,
43]. Anthropomorphic phantoms for dosimetry in diagnostic radiology provide an
advantage over computational simulations in that exact scan geometry and beam
composition do not need to be modeled. However, these phantoms are often
prohibitively expensive. An investigation into low-cost anthropomorphic phantoms was
done at the University of Florida that utilized epoxy resin that forms a hard thermoset
polymer to mimic different human tissues within the phantoms [41]. Soft tissue and lung
tissue equivalent materials used a pliable formulation of this polymer. This particular
phantom was constructed using laser engraving on a high density foam, a technique that
is widely available in university machine shops. These phantoms were constructed in 5-
mm thick slices from head to pelvis and modeled off of DICOM data.
10
Similar studies have been done in creating flexible tissue-equivalent phantoms
for dosimetry and evaluation of diagnostic imaging procedures that would be difficult
to model in software, such as tube current modulation and other proprietary software
mechanisms [42]. The International Committee on Radiation Units has recognized that
affordable tissue-equivalent anthropomorphic phantoms are necessary and thus has
released recommendations on tissue equivalent materials based on elemental
composition, effective atomic number, density and attenuation coefficient [43]. While
these anthropomorphic phantoms create extremely realistic radiographic or computed
tomography images, for the purposes of CACSSI characterization, simply mimicking the
attenuative properties of tissue is not sufficient. The basis of CACSSI is to utilize the
coherent scatter properties of tissues, and therefore our phantoms must be scatter
equivalent in addition to attenuation equivalent. A material with similar attenuation
properties but dissimilar scatter properties to human tissues would not be useful in
CACSSI. Our focus was on finding scatter-equivalent tissue surrogates that could be
used to generate realistic attenuation- and scatter-equivalent phantoms for use in
CACSSI.
11
2. Phantom Development The long-term goals of the CACSSI system in breast imaging are to be used in
intra-operative margin detection, as a complement to mammography or even as an
occasional replacement for breast biopsy. Thus, it is crucial that the system be tested on
human geometry with human materials to assess its capability of functioning in an in-
vivo scan environment. To say that CACSSI can identify a container of graphite within
the beam is not enough. Phantoms were made in human geometries to determine
whether a scanned sample could be detected within a large, dense geometry such as the
human breast.
Phantoms were created in order to characterize the abilities of the CACSSI
system with the aim of accurately modeling human geometries and tissue-scatter
properties within those phantoms. Phantom development began with a calibration
phase using small scan geometries. The calibration materials were scanned alone and
then inside a larger phantom filled with lard. This was done by comparing both
reconstructed form factors to ensure that the material could be accurately classified
within a large phantom geometry.
A phantom that accurately mimicked human anatomical structure was desired.
Actual patient CT data as well as XCAT phantoms [37] were manipulated in order to 3D
print exact replicas of the outer structure of the human anatomy from the DICOM data.
These models could easily be scaled for different size phantoms and re-printed. The
12
relative ease of access to 3D printing in mass quantities was a motivator behind this
approach.
2.1 Calibration Phantoms
Calibration phantoms were derived from previous work using CACSSI in a DHS
experiment. These phantoms typically consisted of a powdered metal, such as graphite
or aluminum, with a strong scatter signal to act almost as a ground truth. These powders
were placed inside a small Nalgene container and placed in the beam path. These were
used to ensure the comparison between form factors of a material scanned by itself and
then again within the phantom. Water was also used as a calibration material as it had a
unique scatter signature compared to any tissue or tissue-equivalent materials.
Figure 3: Powdered metals were inserted into small Nalgene containers as calibration phantoms.
2.2 Scatter Equivalent Materials as Tissue Surrogates
The aim was to find reusable and non-perishable materials with similar scatter
signatures to those found in human breast tissue for use in the phantom. The key is that
the materials need not be equivalent to human tissue in terms of density or effective
atomic number, but becomes challenging when using x-ray diffraction to compare
scatter signatures. The non-perishable nature of these materials would eliminate the
13
need for refrigeration, freezing and thawing for each scan. The most important factor in
assessing similarity of phantom materials and human tissue is the peak location of their
scatter signature. Four reference form factors were targeted for the breast cancer studies:
normal, malignant, benign and fibroglandular.
Food-grade lard (Armour brand) was used as a surrogate for adipose tissue in
the breast. We sought a muscle-type meat to be used in lieu of fibroglandular tissue, and
dried beef was a close surrogate that did not need refrigeration. The form factor for
dried beef resembled that of fibroglandular tissue. Form factors for the PLA used for
printing were also obtained with X-ray diffraction and loaded into our form factor
library.
Table 1: A description of materials used within the phantom and the purpose they serve.
14
Figure 4: A comparison of form factors from human tissues (dashed) and their phantom surrogates (solid). Of note is the similarity of peak location and spectral
shape of the human materials and their surrogates. [34]
Figure 4: Calibration material form factors of graphite (red) and aluminum (blue). These non-biological materials were used in initial testing of the phantoms to
ensure that differences in tissue could be identified within the phantom.
15
Table 2: This table shows the correlation between the human tissue and the tissue surrogates used within the phantom.
A scatter equivalent printing filament was not actively sought. However,
investigation into the tissue equivalence of the printing inks in terms of scatter
signatures was conducted. It was found that the plastic filament used for 3D printing
had a high correlation with lard used to mimic human fat and thus was classified as lard
within the classification maps. These results are shown in figure 5.
The correlations between all of the different materials used within the phantom
as well as the tissues they were meant to mimic were calculated. These correlation
calculations were performed using the known form factors in our library obtained from
X-ray diffraction. Table 3 shows these correlation values. The CACSSI system uses a
threshold of 0.8 for classification and uses the highest correlation value above this
16
threshold for classification. This table shows that the correlation between the calibration
phantoms, aluminum and graphite, are significantly low and thus would never be
confused in a classification. Of particular interest is the correlation between human
tissues. For instance, the correlation between adipose and cancer is .948 based on form
factor shape. Thus, unless normal tissue or something more highly correlated with
adipose tissue were also present in that reconstructed voxel, adipose would be wrongly
classified as cancer by CACSSI. However, by manually visualizing the form factors and
calculating the correlation between the reconstruction form factor and those for adipose
and cancer, the user can determine which material the reconstruction is more closely
correlated with. The value of the correlation between human tissue materials are thus
not necessarily meaningful without the presence of a reconstructed form factor to
compare them to. Form factors differ from each other not only in shape, but in peak
height and location (when they are not normalized). The correlation calculation used in
this work considers only the overall form factor shape. Previous work was done by
Lakshmanan [8] in analyzing the q-value at full peak height normalized by the q-value
at half intensity. Further work could be done to make the CACSSI algorithm more
robust by incorporating such parameters for peak height location and relative height
into the reconstruction code.
17
Figure 5: A comparison of the scatter signatures from the 3D printing filament (blue), lard (green) and reconstructed lard (thin blue). The correlation between lard
and plastic form factors is 0 .953, meaning the plastic would be classified as lard within the CACSSI reconstruction.
18
Table 3: A table of correlation values between materials used in the phantom as well as the tissues they mimicked.
Materials Correlation Adipose Aluminum 0.338 Adipose Cancer 0.948 Adipose Fibroglandular 0.981 Adipose Graphite 0.309 Adipose Normal 0.987
Aluminum Cancer 0.392 Aluminum Fibroglandular 0.361 Aluminum Graphite 0.091 Aluminum Lard 0.214 Aluminum Normal 0.360
Cancer Fibroglandular 0.972 Cancer Graphite 0.365 Cancer Lard 0.737 Cancer Normal 0.967
Fibroglandular Graphite 0.341 Fibroglandular Lard 0.782 Fibroglandular Normal 0.996
Graphite Lard 0.237 Graphite Normal 0.334
Lard Normal 0.805
19
Table 4: Correlations of different materials with the adipose form factor. Lard is the one chosen as the surrogate material because it has the highest correlation.
Table 5: Correlation of different materials with the cancer form factor. Dried Beef is the one chosen as the surrogate material because it has the highest correlation,
although excised human tumors were eventually used in the phantoms.
20
Table 6: Correlation of different materials with the fibroglandular form factor. Dried beef is the one chosen as the surrogate material because it has the highest
correlation as the dried beef was phased out as the cancer surrogate when human tissue samples were acquired.
Table 7: Correlation of different materials with the normal tissue form factor. Normal human tissue is classified as an even mixture between fibroglandular and
adipose tissues and thus lard and dried beef sufficed.
21
2.1.1 Previous Work
In Kidane [3], form factors were measured using energy dispersive X-ray
diffraction (EDXRD). The sample was rotated during scan acquisition. Tissue samples
were obtained from 100 mastectomies, lumpectomies and breast reductions over a very
wide age range. Tissues identified consisted of fibroglandular (50% fibrous, 50%
glandular), adipose and healthy (a mixture of fibroglandular and adipose) [3]. The
relative presence of these tissue types in each sample was determined through histology.
Carcinoma was also present. According to Kidane, the presence of micro calcifications
did not affect form factor measurements significantly. The samples were snap frozen
and placed in a thin-walled sample holder for scanning. The thin walls of the sample
holder minimized attenuation and are similarly used in calibration of our system. The
scanned specimen was brought back to room temperature for the measurement and it
was found that the freezing and thawing process did not affect the form factors
measured, which we assumed for our own work.
According to Kidane, the factors that influence measured form factors include
beam hardening, the shape of the incident X-ray spectrum, Compton and multiple
scatter in the sample, physical scatter from the imaging system, and the fluctuation of
the incident spectra [3]. Since a few of these factors were found in Kidane to be
insignificant, the recorded photon count was corrected by incident spectrum shape,
background scatter and the energy dependence of attenuation. Energy measurements
22
were transformed into a function of momentum transfer and background radiation was
subtracted from the measured spectra.
2.3 Early Phantom Models
The first generation of the breast phantom was a circular Tupperware container
filled with lard and dried beef. The prototype for the 3D printed phantom was a
cylindrical container with Nalgene-sized holes in which different materials could be
inserted in any combination. The body of the phantom could be filled with air, water or
lard depending on preference. The second generation of the breast phantom was a 3D
printed, hollowed semi-sphere on a flat base. This phantom was designed in SketchUp
with removable shelves placed inside. The shelves could be used to easily place tissue
samples or calibration materials at known heights within the phantom geometry.
Figure 6: This version of the phantom included “shelves” for the convenient placement of excised human tissue samples or calibration phantoms.
23
2.4 Current Phantom Models
The third and most current generation of the breast phantom was obtained from
DICOM data supplied by the Duke MMIL lab (courtesy Martin Tornai). This phantom
was based on a prone scan of a breast on a dedicated breast CT system. This anatomical
orientation was chosen partly because it was a way to incorporate the excised human
tissue specimens in a natural geometry, rather than having to imitate a compressed state
when incorporating the excised tissue into the phantom. It was also a readily available
data set that we could compare our phantom to by reproducing the scan within the same
CT unit. The effective diameter of the phantom is 8.3 cm. This is approximately in
between the typical compressed breast diameter of 5.2 cm and the average human breast
diameter of 13.1 cm. It is envisioned that CACSSI could be used for prone breast
geometries in order to minimize scatter coming from compression paddles that would
interfere with the scatter measurements from the tissue.
Figure 7: A SketchUp rendering of a DICOM modeled prone breast phantom
24
Figure 8: The 3D printed model of the rendering in figure 1.
Figure 9: A mammogram of the phantom in the horizontal plane shows the hexagonal infill structure of the phantom shell, which presents itself as vertical lines
in a coronal or sagittal view.
2.4.1 Prone Breast Geometry
Prone breast setup was historically used for patients with large and pendulous
breasts to maximize the distance between the treatment volume and the heart and lung
for radiation treatments. The NYU Langone Medical Center [29] has shown that treating
patients in a prone setup can be beneficial in reducing toxicities, especially to vital
25
organs like the heart and lung. The use of a prone scanning protocol also helps to reduce
respiratory motion that would affect scan acquisition. This setup is typically done using
a breast board with arm supports for reproducibility in both diagnostic CT and radiation
therapy treatments. However, the board tends to introduce artifacts, and several other
considerations must be made in terms of couch and patient positioning, light field
obstruction, ease of use and patient comfort. While the study at NYU demonstrated
reproducibility in several factors of patient positioning, the dedicated breast CT
developed at Duke is a solution to many of the positioning parameters to be considered
in using conventional CT in breast imaging. We used prone breast DICOM data for our
phantoms due to the reliable positioning reproducibility, as well as the availability of the
dedicated breast CT in a collaborating lab.
2.4.2 The 3D Printing Process
The advent of additive manufacturing has been pivotal in the field of biomedical
research and clinical practice. Applications include fabrication of tissues, organs,
prosthetics and implants [30]. 3D printing can contribute to customized medicine in
terms of being able to print to the needs of each individual patient. There is also the
ability to print specialized fixtures used in clinical settings like operating and exam
rooms. The relative ease and low-cost of 3D printing, combined with the short time
required for printing, sets the stage for a revolution in medicine that is often compared
to the way in which the printing press revolutionized publishing [30]
26
In determining how to best physically model any geometry to our exact
specifications, we took advantage of the vast 3D printing resources available at Duke.
The Duke Innovation Studio houses 33 3D printers that can be operated remotely using
a cloud-based platform called 3DPrinterOS. This software uploads a .stl file, a format
from stereolithography CAD software, which is a surface “mesh” rendering of the object
to be printed. This mesh is a series of linked triangles. An .stl mesh is imported from a
computer aided design (CAD) software such as SketchUp and ITK-Snap [44], both of
which were used in this work and will be discussed in detail later. This .stl file can be
manipulated in terms of size, scale and proportion within 3DPrinterOS in a CAD-like
setting. Often, there is some “damage” inherent in the .stl file that is created. This is
because the geometries to be printed are from scanned object data and are not built in a
series of linked triangles. A service used by 3DPrinterOS, called Netfabb, is used to fix
these damages by making the mesh into a solid object without holes or incorrectly
oriented surfaces (NETFABB). This mesh is then “sliced” within the 3DPrinterOS
software in the conversion to a .gcode format. The slicer “cuts” the 3D model into 2D
slices. Each slice height is the same dimension as the extruder filament [31]. The slicing
is analogous to a computer aided manufacturing (CAM) process in traditional
computerized numeric control (CNC) machining. Slicing takes into account several
printer parameters including, but not limited to, nozzle diameter, melting point of the
27
plastic ink and extrusion speed [31]. Thus, the .gcode file defines the parameters used to
print the .stl file as designed.
3DPrinterOS also acts as a “host” for user interfacing with the printer. Thus, files
can be uploaded, manipulated, sliced and sent to the printer remotely. In a process
similar to a “FileàPrint”, the user selects the printing speed, percent of infill, layer
thickness and several other parameters. Decreasing the layer thickness and increasing
the infill can make a scan time prohibitively long (on the order of several days) and are
generally unnecessary for the purposes of this work. Typically, a print is done with a
shell thickness of .4-mm, layer thickness of .15-mm and an infill of 10%. The temperature
of the printer extruder is 240°C and the printing bed is 70°C. Supports are often included
on a print since overhang angles of greater than 45° are not easily printed. These
supports are loosely attached to the body of the printed structure and are easily
removed after printing. A helpful feature of the Innovation Studio is the use of cameras
on each printer to monitor the object as it is being printed. This gave us the opportunity
to visualize the inner structure of the phantom shell as it was being printed.
The printers used in this work consist of a Makerbot Replicator Desktop 3D
Printer and an Ultimaker 2. The Makerbot was used to print all of the original breast
phantoms used in this work and the Ultimaker was used for experimental prints of other
organs, as well as extra prints of the breast phantom. These printers are Fused
Deposition Modeling (FDM) printers that use thermoplastic inks like Acylonitrile
28
Butadiene Styrene (ABS) and Poly Lactic Acid (PLA). These materials come in spools
and are fed into the extruder, become malleable when heated, and return to a solid upon
cooling. ABS is an oil-based thermoplastic that requires a heated bed to minimize
warping that occurs when it is cooled during printing. PLA is made from cornstarch and
sugarcane and is generally a weaker material than ABS because it is prone to cracks and
warping. However, for the purposes of this work, PLA was used as it was readily
available.
2.4.3 Segmentation of DICOM Data
The experience with imaging the breast phantom lead to an interest in other
organs, such as lung and spine. In these cases, XCAT phantom data was obtained from
the Segars lab[37]. The DICOM images were viewed first in ImageJ, an image processing
program developed at the National Institutes of Health. After determining which data
sets we wanted to print, these files were opened in ITK-Snap, a software founded as a
collaboration between the University of Pennsylvania and the University of Utah that
facilitates delineation of anatomical borders and automatic image segmentation based
on those borders. These scans were skull-base to mid-thigh (SBMT) protocols that
allowed for easy delineation of bone and air from the other soft tissue present in the
body. There are several methods that can be used to define separate anatomical regions
of interest.
29
ITK-Snap uses an active contour, or “snake”, for semi-automatic active
contouring. After selecting a region of interest around the organ I plan to segment, there
is a pre-segmentation step done using thresholding, classification, edge attraction or
clustering. To contour an organ like the lung or bone, which has inherently high contrast
with other soft tissues in the body, I use a classification pre-segmentation. This involves
placing colored labels in different tissue types in order to train the classifier. Small
regions of interest are then placed within the organ of interest and expand throughout
the organ as determined by the classifier. This can then be saved as a surface mesh and
printed.
The breast phantom was created using the thresholding method, where an upper
and lower grey value threshold are set to differentiate between soft tissue. Since the
contrast of different breast tissues is low, the classification algorithm cannot be trained
to differentiate by such a subtle amount. The thresholding method can be used to finely
tune the differences in grey levels of each soft tissue to separate the breast tissue from
the chest wall and other tissues inherent in the thoracic region.
30
Figure 10: An image of the ITK-Snap interface while segmenting the lungs using the classification trainer.
Figure 11: The printed phantom from the segmentation shown in figure 10.
31
3. Phantom Validation and Verification The first step in this work was to be sure that our system could identify and
resolve different materials present in the beam. Calibration phantoms with strong scatter
signals, such as aluminum and graphite powders, were used for initial testing of the
system. The same materials were then surrounded with lard and placed in the phantom
and scanned again. The purpose of this process was to ensure that the system could
identify the materials alone as well as within the phantom geometries.
3.1 Validation of Anatomy
A mammogram was taken of the breast phantom with both cancer and normal
tissue samples placed in it with lard. The scan was taken on a Hologic Selenia
Dimensions mammographic unit in the Duke Cancer Center. The scan protocol was a
Tomosynthesis image with 28 kVp and 100 mA for 272 microseconds using a 0.3-mm
focal spot size. A Tungsten target and Rhodium filter was used as determined by the
unit. The compression paddle was used to hold the phantom in place but did not exert
any force on the phantom.
A helical CT was also obtained on a GE LightSpeed Xtra Computed Tomography
unit in the Duke Hospital. The scan protocol used a 1-mm slice thickness, 80 kVp, 100
mA and 1025 microseconds and reconstructed using a standard convolution kernel and
a 512 x 512 matrix. The phantom was scanned with lard alone followed by a Nalgene
filled with graphite placed inside the phantom.
32
Table 3 shows the Hounsfield number scale for different tissues, as well as the
Hounsfield numbers for the lard and plastic filament used in the phantom.
Figure 12: The setup used to take a mammogram of the phantom.
Figure 13: The resulting mammographic image. The “streaks” in the image are due to the plastic infill within the phantom shell.
33
Figure 14: A CT scan of the phantom filled with lard. The phantom on the right also contains a Nalgene vial filled with graphite.
Table 8: A table of the Hounsfield Numbers of different materials
Material CT Number
Lard -147
Plastic Filament -5
Air -1000
Water 0
Lung -500
Soft Tissue 100 to 300
Muscle 10 to 40
Fat -100 to -50
34
3.2 Form Factors
Scatter signature measurements of calibration materials for the system were
made using an X-ray diffraction system with a Panalytical XPert PRO HR diffraction
system with a 1.8-kW sealed ceramic X-ray source and a Ni .125-mm automatic beam
attenuator [7]. These calibration materials included several metallic powders, such as
aluminum sulfate, graphite and sodium chloride. The calibration powders and liquids
are housed in 10-mm diameter Nalgene containers made of low density polyethylene
[6]. These, and other materials were scanned and tested within our group for verification
and validation of the imaging system [35].
3.3 Scanning Process
The experimental system comprises a Varian G1593BI X-ray tube with rotating
Tungsten-rhenium anode and .8mm focal spot, several collimators, a coded aperture and
an energy sensitive 128-pixel linear array detector. The linear array detector, made of
CdTe with .8mm2 pixels, makes the system more compact than an area detector setup.
The detector array is placed perpendicular to the beam out of the primary beam path.
Over a range of 20-170 keV, the energy resolution is nearly uniform with FWHM=6 keV
[6]. The first generation of the system used for scanning phantoms used a series of lead
collimators to reduce the beam to a pencil with an angular divergence of ~1 mrad. The
coded aperture enables resolution along the beam direction by modulating the scattered
X-rays. This eliminates the need to rotate the sample and allows for volumetric
35
reconstruction from a 2D raster scan [8] The coded aperture was machined into 1-mm
thick slits from a 1-mm slab of bismuth-tin alloy. Later generations of the system used a
Ralco collimator with light field and laser alignment to collimate the beam as would be
done on a clinical tube. The coded aperture was replaced with one that looked like a QR
code as shown in figure 2. The tube was operated at 125 kVp and 40 mA for 10 seconds
per scan, or at 125 kVp and 10 mA for 3-5 minutes using fluoroscopy mode.
Conventional radiography has inherently low detection efficiency as only ~<1% of
primary X-rays typically reach the detector due to attenuation within the patient.
Coherent scatter typically only accounts for a few percent of X-rays reaching the
detector, and therefore to have any measureable scatter signal requires a very high flux
of electrons from the tube. We typically measure less than 1% of the total coherent
scatter due to the very small detector cross section employed in the CACSSI system [6].
This is the reason for using such a long scan time (10s for snapshot mode and ~minutes
for fluoroscopy mode) as compared to conventional clinical procedures.
36
Figure 15: The scanning geometry. A is the x-ray tube, B is a collimator, C is the object stage, D is the coded aperture and E is the detector. The samples are scanned
using 125 kVp and 100 mAs with a ~1mm collimated pencil beam.. [34,35]
The tissue samples initially used had been repeatedly frozen and thawed for
several years. This may or may not have had an effect on the structural integrity of the
tissue, which would impact the coherent scatter signature obtained from it. Two
samples were matched from the same patient- one was cancerous and one was healthy.
The samples were about 2.5-cm in length and width and weight approximately 40g.
These samples were raster scanned in the system with 1-mm transverse resolution to
develop the classification map shown. Approximately 260 total measurements were
taken for the raster scan.
37
3.3.1 Raw Data
The raw data from each scan acquisition is a .bin file that is transformed into a
.mat file in the first set of code. In this code, the user specifies whether the scan was
background or signal and saves a .mat file accordingly. This .mat file is the signal
summed over all time frames and includes the background. A bell shaped curve is
generated which shows the coherent scatter signal from the sample that comes through
the coded aperture. The bell shape results from lower energy scatter occurring at larger
scattering angles [39].
Figure 16: An example of the bell shaped curve generated in the raw data with y-axis representing Energy [keV] and x-axis representing pixel number.
The peak of this bell curve indicates the position of the scanned sample along the
x-axis. The energy of the signal measured is inversely proportional to scatter angle.
Thus, a higher energy measured signal reflects a lower scatter angle. However, the
probability of coherent interaction is greater at lower energies. Thus, at lower energies,
the cross section is larger and more scatter is inherent on the detector.
38
3.3.2 Reconstruction
The reconstruction works on the basis that the origin of the scatter is at a known
location. Therefore, a forward matrix is calibrated for an object at a known location with
a beam of a known width. Using the sample location along with the coded aperture
allows for reconstruction of the object in three dimensions.
All of the scatter data acquired was reconstructed using the maximum a
posteriori (MAP) estimation and total variation (TV) CACSSI algorithm and was then
classified as being a particular breast tissue type by matching its form factor to the
ground-truth form factors. The TV smoothes the data while preserving edges and the
data is then thresholded to generate a classification map. The threshold is set based on a
percentage correlation to a known form factor. Typically, this threshold is set at
approximately 85% such that a scanned material will only be classified as a material
from the material library if they have an 85% correlation to the known form factor. The
number of MAP iterations used is specified by the user. It was found that a low number
of iterations (<3) was good for reconstructing the form factors without much noise, but a
larger number of iterations (~50) yields better spatial resolution of the reconstructed
classification map. We often reconstructed the data twice using each set of iterations
such that we got the clearest data for the task. Voxels containing more than one tissue
type exhibited the peaks from both tissue types in their form factors and were therefore
manually classified as mixtures of those tissues. Several years ago, these tumors had
39
routine histological processing, which involved dehydration, clearing, infiltration and
being embedded in paraffin wax. They were sectioned with a microtome and stained
with H&E staining to be interpreted by a pathologist [8]. The classification images and
pathological maps were registered based on visible edge and geometric features.
Figure 17: This is the first plot shown in the reconstruction code. This example was taken from a pencil beam scan of lard alone. The user clicks the spot on the z-axis where the object was known to be placed (which usually corresponds to the brightest
white pixel in this image). That gives the reconstructed form factors through that pixel.
Figure 18: This figure shows the distribution of tissue types along the beam axis from the same scan as figure 15 above. The color map indicates that all of the q
values detected by the system matched that of lard.
Noise in this system comes from several factors including detector imperfections.
Dead pixels are not counted and thus contribute no information and are indicated by
40
vertical black lines along the raw data images. There is also the presence of photons at
different energies originating from other scatter mechanisms hitting the detector,
including Compton scatter. Large objects also get blurred in the range direction due to
the ~1cm depth resolution. There is also the presence of “side lobes” in the raw data that
are inherent in the coded aperture. As long as we know these factors are present, we can
correct for them in the reconstruction process.
3.4 Verification Methodology
We have vacuum sealed tissues will for transportation and scanning in our
system. We first had to ensure that the vacuum bags themselves would not introduce
any artifacts into our images, as well as make sure they had a form factor that was
different from the materials we hope to detect inside of it. To do this, we scanned a
background, a 5-bag thick sample, a piece of chicken breast alone, and the same chicken
sample inside the vacuum bags. A visual analysis of the raw data showed that the bag
has some degree of self-attenuation when placed several layers thick within the beam,
but has a signature that is different from the chicken breast. The chicken breast sample is
negligibly different from the human adipose tissue we use in terms of scatter signatures.
41
Figure 19: Raw data from scans taken of (a) background signal, (b) the vacuum bags, (c) chicken breast and (d) chicken breast inside the vacuum bags. The decrease
in counts from (a) to (b) shows self attenuation of the bags.
Figure 20: This figure shows the subtraction of the chicken alone data (18.c) and the chicken in the vacuum bag (18.d). Rather than indicating a strong scatter
pattern, this subtracted image shows random noise that indicates the bags do not have a strong signal in the scatter imaging system.
42
Table 9: A table of materials scanned and metrics of comparison used in analysis. The materials were first scanned alone in the beam, and then surrounded by lard within the phantom. Several of the metrics of comparison were simply visual: number of major peaks, overall peak shape and peak shift. Peak shift indicates a shift of the
material form factor toward the peak of lard when placed inside the phantom. The peak location along the x-axis is both a visual and quantitative value. The correlation of the material alone and within the phantom is a quantitative metric described in a
later section.
43
4. Results and Discussion There are several important parameters in a reconstructed form factor that allow
tissue identification. They include peak location along the momentum transfer axis,
number of peaks and general shape of the form factor.
4.1 Metrics of Comparison
Visual comparisons of form factors are done to count the number of peaks and to
assess the similarity in shape (including concavity) of reconstructed form factors and
known form factors. Most form factors have one major peak along the momentum
transfer axis. However, some materials (such as lard) have smaller intensity peaks that
can be used as an identifier. For example, lard and adipose have the same major peak
location but lard has secondary and tertiary peaks that can be used to differentiate lard
form factors from adipose. The location of the major peak is also indicative of the
material present in the beam. It was typically seen that the form factor of an object
within the lard in the phantom often has the same shape as the form factor of that object
alone, but with a peak shifted toward the location of the peak of the lard spectrum.
To quantitatively assess the similarity in shape between two unique form factors,
a correlation was done between the two spectra normalized by their 𝚤. (Euclidean)-
norm. The 𝚤. – norm of a vector v is defined as:
𝚤. = 𝑣 ⋅ 𝑣
44
The correlation between unique spectra a and b is computed by:
𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 = (𝑎/𝑙<.)⋅ (𝑏/𝑙>.)
where 𝑙?. is the 𝚤.-norm of that spectrum. By the Cauchy-Schwarz Inequality, a
correlation calculation done between two identical vectors should yield a correlation of
1. Thus, the closer the numerical value of the correlation to 1, the more similar they are.
Previous CACSSI experiments have set the precedent of using a threshold correlation of
0.8 to identify as a material spectrum as correlating enough with a spectrum from the
library that it can be classified as that material. However, experimentation of this
threshold in past works have shown that setting the threshold anywhere within the
range of .7 to .95 is sufficient for tissue classification in CACSSI.
Figure 21: Reconstructed form factor of lard within the phantom geometry (blue) and lard alone (solid green). The correlation is .983.
45
Table 10: Lard metric results
Figure 22: Reconstructed form factors of dried beef alone (left) and dried beef inside the breast phantom filled with lard (right). Of note is the dual peaks on the
right which indicate the presence of both lard and beef.
46
Figure 23: Extracted reconstructed form factors of beef alone (blue) and beef inside lard in the phantom (red) plotted together for comparison. Their correlation is
.911.
Table 11: Beef metric results.
47
Figure 24: As mentioned in section 1.2.1, it was determined that the order of materials placed within the beam does not affect their ability to be detected. The image on the left is the form factor of beef followed by lard, and the image on the
right is the form factor of lard followed by beef within the beam.
Figure 25: Extracted reconstructed form factors of beef followed by lard (blue) and lard followed by beef in the phantom (red) plotted together for comparison. Their
correlation is .925 and shows that the signal differs mostly in amplitude rather than shape or peak location. This demonstrates that the ordering of objects within the
beam does not affect their ability to identified.
48
Figure 26: Reconstructed form factors of water alone (left) and water inside the breast phantom filled with lard (right). Water was used as a calibration material. Of note is the shift on the right of the reconstructed form factor (blue) toward the lard
form factor, which indicate the presence of both lard and water in the beam.
Figure 27: Extracted reconstructed form factors of water alone (blue) and water inside lard in the phantom (red) plotted together for comparison. Their correlation is
.891.
49
Table 12: Water metric results.
Figure 28: Reconstructed form factors of tumor alone (left) and tumor in lard inside the phantom (right). Of note is the change in reconstructed form factor shape between the two plots, the second reconstruction more closely mimicking the shape
of the lard form factor.
50
Figure 29: Extracted reconstructed form factors of tumor alone (blue) and tumor inside lard in the phantom (red) plotted together for comparison. Their
correlation is .985.
Table 13: Tumor metric results.
51
Figure 30: Reconstructed form factors of normal tissue alone (left) and normal tissue in lard inside the phantom (right). Of note is the change in reconstructed form
factor shape between the two plots, the second reconstruction more closely mimicking the shape of the lard form factor.
Figure 31: Extracted reconstructed form factors of normal tissue alone (red) and normal tissue inside lard in the phantom (blue) plotted together for comparison.
Their correlation is .885.
52
Table 14: Normal metric results.
Table 15: This table shows the correlation between the form factor of a material with the form factor of the same material placed in the phantom with lard.
While the threshold for classification is typically set to a correlation of 0.8,
thresholds as low as .7 have been shown to yield reliable results. The low correlation of
53
tumor alone with tumor inside the phantom could thus be classified as tumor in the
classification map if the threshold in the CACSSI reconstruction algorithm is lowered.
Figure 28 below shows the classification map of the tissue sample used for the tumor
scans. The low correlation between tumor alone and tumor inside the phantom is likely
due to the tumor cross-section of the tissue sample being so small relative to that of
normal tissue.
This work shows that CACSSI is able to classify a material both when it is
scanned alone as well as when it is placed inside lard within the phantom geometry.
CACSSI’s capability to discern small volumes of material within a relatively large
volume of phantom geometry is significant because the same ability will likely translate
well to in-vivo applications. Detecting a small volume of material within a lard-filled
vessel is analogous to detecting a small lesion within a dense human breast in terms of
coherent scatter imaging. Future phantom work will include building a human torso-
equivalent phantom to assess the capability of CACSSI to classify tissues within several
different anatomical structures. This will afford the opportunity to investigate different
cancer pathologies (lung, thyroid, liver) within the CACSSI system.
54
Figure 32: (a) A tumor classification map rendered by CACSSI. (b) An interpolated image to register with the histological report in (c). [35] This is an
example of how the form factor from each voxel is matched to a spectrum from the form factor library through correlation and classified as a material.
55
5. Conclusions and Future Work It has been shown that through both visual and qualitative metrics, a user can
identify the presence of a material within the depth of the phantom based on form factor
analysis. This is important in the translation of CACSSI to an ex-vivo modality to
something that might be used in-vivo for diagnostic purposes. The short-term goal is to
use CACSSI for intra-operative margin detection. This work has shown that the system
is able to identify different human tissue types in accurate human geometries. Future
work will include
5.1 Pathology Protocol
A new IRB protocol has been established to intercept fresh biopsied tissue
between the operating room and pathology. After excision, the tissue sample will be
vacuum-sealed and scanned in our system. It will then be sent back to pathology where
regular histological processes are done. The pathology report will come back to us so
that we can determine if our classification map rendered in the system matches what
was found in pathology. The nominal time frame for this process is about 30 minutes
until the sample is sent to pathology for final analysis. Refinement of the protocol is
needed to decrease the total time. The limiting factor in terms of the time required for
the entire process is the scan time. Scan parameters such as beam width, translation
stage speed and conventional radiography versus fluoroscopic mode will all have a part
to play in reducing the procedural time. A shorter time, on the order of 10 minutes,
56
would mimic the desired timeline for intra-operative margin detection as discussed in
previous works.
5.2 Printing Filament Experimentation
Flexible polymers for use in anthropomorphic phantoms were discussed in an
earlier section of this work. A future plan is to investigate the scatter properties, and
general performance, of different 3D printable plastic filaments. Of particular interest is
a pliable filament called NinjaFlex made from a thermoplastic polyurethane, which is
more durable and resistant to wear than its ABS or PLA competitors [40]. The filament,
when printed, is rubber-like in texture. The infill percentage can be manipulated such
that with a small infill, the printed product is extremely flexible. When there is a very
high degree of infill, the printed object texture and flexibility resembles that of a car tire.
The NinjaFlex has a very low resistance exterior and a uniform filament diameter
compared to other commercially available filaments which allows for a very smooth and
uniform print. In addition, the NinjaFlex formulation is resistant to many industrial
chemicals such as petroleum and Freon and is often used to print gaskets and seals [40].
This is an important feature which allows placement of nearly any material (solid or
liquid) within a printed phantom geometry. The printing guidelines in terms of extruder
and bed temperature are the same as PLA and ABS and thus can be printed using the
same printers.
57
5.3 Solid 3D Printed Phantom
There is work being done within the Ravin Advanced Imaging Labs on
customized 3D printing filament doped with different elements or chemicals.
Investigation into the scatter properties of these inks will be done and compared to the
human tissues of interest by a similar correlation method that was utilized in this work.
A “match” for each breast tissue type would ideally be created. A solid 3D printed
phantom could then be created using the matching filaments in a known geometry.
Structures such as microcalcifications and milk ducts could also be introduced into the
phantom geometry. Such a phantom could then be used as a ground truth measurement
for characterization of CACSSI and would eliminate the need for preservation of human
tissue, as well as filling and emptying a hollow phantom shell with scatter-equivalent
materials.
58
References
1. American Cancer Society. (2013, Aug) Breast cancer.
2. Johns P. Coherent scatter in diagnostic radiology. Med Phys. 1983;10(1):40.
3. Kidane G, Speller R, Royle G, Hanby A. X-ray scatter signatures for normal and neoplastic breast tissues. Physics in Medicine and Biology. 1999;44(7):1791-1802.
4. W. Friedrich, P. Knipping, and M. v. Laue, “Interference appearances in xrays,” Ann. Phys.(Berlin), vol. 41, pp. 971–988, 1913
5. W. Bragg and W. Bragg, “The reflection of X-rays by crystals,” Proceedings of the
Royal Society of London. Series A, vol. 88, no. 605, pp. 428–438, 1913.
6. Greenberg J, Krishnamurthy K, Brady D. Snapshot molecular imaging using coded energy-sensitive detection. Opt Express. 2013;21(21):25480.
7. Lakshmanan M, Kapadia A, Sahbaee P, Wolter S, Harrawood B, Brady D et al. An X-ray scatter system for material identification in cluttered objects: A Monte Carlo simulation study. Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms. 2014;335:31-38.
8. Lakshmanan M. X-ray Coherent Scatter Imaging for Intra-Operative Margin Detection in Breast Conserving Surgeries. Duke University, 2015
9. C. E. DeSantis, C. C. Lin, A. B. Mariotto, R. L. Siegel, K. D. Stein, J. L. Kramer, R.
Alteri, A. S. Robbins, and A. Jemal, “Cancer treatment and survivorship statistics, 2014,” CA: a cancer journal for clinicians, vol. 64, no. 4, pp. 252–271, 2014
10. D. M. Cunha, O. R. Oliveira, C. A. Perez, and M. E. Poletti, “X-ray scatteringprofiles of some normal and malignant human breast tissues,” X Ray Spectrom, vol. 35, no. 6, pp. 370–374, Nov 2006.
11. D. L. Batchelar and I. A. Cunningham, “Material-specific analysis using coherent-
scatter imaging,” Medical Physics, vol. 29, no. 8, p. 1651, 2002.
12. Moon H, Kim N, Jeong S, Lee M, Moon H, Kim J et al. The Clinical Significance and Molecular Features of the Spatial Tumor Shapes in Breast Cancers. PLOS ONE. 2015;10(12):e0143811.
59
13. Understanding Triple Negative Breast Cancer [Internet]. Triple Negative Breast
Cancer Foundation. 2016 [cited 17 March 2016]. Available from: http://www.tnbcfoundation.org/understanding-triple-negative-breast-cancer/
14. C. R. F. Castro, R. C. Barroso, and R. T. Lopes, “Scattering signatures for some human tissues using synchrotron radiation,” X Ray Spectrom, vol. 34, no. 6, pp. 477–480, Nov 2005.
15. M. Poletti, O. Goncalves, and I. Mazzaro, “X-ray scattering from human breast
tissues and breast-equivalent materials,” Phys Med Biol, vol. 47, no. p. 47, 2002.
16. E. Ryan and M. Farquharson, “Angular dispersive X-ray scattering from breast tissue using synchrotron radiation,” Radiat Phys Chem, vol. 71, no. 3, pp. 971–972, 2004.
17. E. a. Ryan and M. J. Farquharson, “Breast tissue classification using X-ray scattering measurements and multivariate data analysis,” Phys Med Biol, vol. 52, no. 22, pp. 6679–6696, nov 2007.
18. A. L. C. Concei¸c˜ao, M. Antoniassi, and M. E. Poletti, “Analysis of breast cancer by small angle X-ray scattering (SAXS).” Analyst, vol. 134, no. 6, pp. 1077–82, jun 2009.
19. W. M. Elshemey, O. S. Desouky, M. M. Fekry, S. M. Talaat, and A. a. Elsayed, “The diagnostic capability of X-ray scattering parameters for the characterization of breast cancer,” Med Phys, vol. 37, no. 8, p. 4257, 2010.
20. S. Pani, E. J. Cook, J. a. Horrocks, J. L. Jones, and R. D. Speller, “Characterization of breast tissue using energy-dispersive X-ray diffraction computed tomography.” Appl Radiat Isot, vol. 68, no. 10, pp. 1980–7, oct 2010.
21. S. Sidhu, G. Falzon, S. a. Hart, J. G. Fox, R. a. Lewis, and K. K. W. Siu, “Classification
of breast tissue using a laboratory system for small-angle X-ray scattering (SAXS).” Phys Med Biol, vol. 56, no. 21, pp. 6779–91, nov 2011.
22. W. M. Elshemey, F. S. Mohamed, and I. M. Khater, “X-ray scattering for the characterization of lyophilized breast tissue samples,” Radiat Phys Chem, vol. 90, pp. 67–72, sep 2013.
23. C. Theodorakou and M. J. Farquharson, “Human soft tissue analysis using X-ray or gamma-ray techniques.” Phys Med Biol, vol. 53, no. 11, pp. R111–49, jun 2008.
60
24. S. Kauppila, F. Stenb¨ack, J. Risteli, A. Jukkola, and L. Risteli, “Aberrant type i and
type iii collagen gene expression in human breast cancer in vivo,” The Journal of pathology, vol. 186, no. 3, pp. 262–268, 1998.
25. I. Pappo, R. Spector, A. Schindel, S. Morgenstern, J. Sandbank, L. T. Leider, S.
Schneebaum, S. Lelcuk, and T. Karni, “Diagnostic performance of a novel device for real-time margin assessment in lumpectomy specimens,” Journal of Surgical Research, vol. 160, no. 2, pp. 277–281, 2010.
26. S. Kennedy, J. Geradts, T. Bydlon, J. Q. Brown, J. Gallagher, M. Junker, W. Barry, N. Ramanujam, and L. Wilke, “Optical breast cancer margin assessment: an observational study of the effects of tissue heterogeneity on optical contrast.” Breast cancer research : BCR, vol. 12, no. 6, p. R91, jan 2010.
27. M. D. Keller, E. Vargis, N. de Matos Granja, R. H.Wilson, M.-A. Mycek, M. C. Kelley,
and A. Mahadevan-Jansen, “Development of a spatially offset Raman spectroscopy probe for breast tumor surgical margin evaluation.” Journal of biomedical optics, vol. 16, no. 7, p. 077006, jul 2011.
28. I. J. Bigio, “Real-time pathology to guide breast surgery: seeing alone is not believing.” Clinical cancer research : an official journal of the American Association for Cancer Research, vol. 18, no. 22, pp. 6083–5, nov 2012.
29. No H. Reproducibility in Prone Breast Cancer Treatments. Presentation presented at; 2014; AAMD 39th Annual Conference, Seattle, Washington
30. Ventola CL. Medical Applications for 3D Printing: Current and Projected Uses. Pharmacy and Therapeutics. 2014;39(10):704-711.
31. Hackney M. 3D Printing 101: Part 4: Software - CNCCookbook CNCCookbook [Internet]. Blog.cnccookbook.com. 2016. Available from: http://blog.cnccookbook.com/2013/05/01/3d-printing-101-part-4-software/?nabm=0
32. Greenberg J, Hassan M, Krishnamurthy K, Brady D. Structured illumination for tomographic X-ray diffraction imaging. The Analyst. 2014;139(4):709-713.
33. Types of Breast Cancer [Internet]. Pathology.jhu.edu. 2016 [cited 15 March 2016]. Available from: http://pathology.jhu.edu/breast/types.php
61
34. Albanese K, Morris R, Lakshmanan M, Greenberg J, Kapadia A. MO-F-CAMPUS-I-03: Tissue Equivalent Material Phantom to Test and Optimize Coherent Scatter Imaging for Tumor Classification. Med Phys. 2015;42(6):3575-3575.
35. Morris R, Albanese K, Lakshmanan M, Greenberg J, Kapadia A. MO-F-CAMPUS-I-04: Characterization of Fan Beam Coded Aperture Coherent Scatter Spectral Imaging Methods for Differentiation of Normal and Neoplastic Breast Structures. Med Phys. 2015;42(6):3575-3575.
36. Lakshmanan M, Greenberg J, Samei E, Kapadia A. Design and implementation of coded aperture coherent scatter spectral imaging of cancerous and healthy breast tissue samples. Journal of Medical Imaging. 2016;3(1):013505.
37. Segars WP, Sturgeon GM, Mendonca S, Grimes J and Tsui BMW. 4D XCAT Phantom
for Multimodality Imaging Research. Medical Physics, vol. 37, pp.4902-4915, 2010. 38. Cui C, Jorgensen S, Eaker D, Ritman E. Direct three-dimensional coherently scattered
x-ray microtomography. Med Phys. 2010;37(12):6317. 39. Pang S, Hassan M, Greenberg J, Holmgren A, Krishnamurthy K, Brady D.
Complementary coded apertures for 4-dimensional x-ray coherent scatter imaging. Opt Express. 2014;22(19):22925.
40. NinjaFlex® Flexible 3D Printing Filament | NinjaTek™ [Internet]. NinjaTek. 2016
[cited 12 April 2016]. Available from: http://www.ninjaflex3d.com/products/ninjaflex-filaments/
41. Winslow J, Hyer D, Fisher R, Tien C, Hintenlang D. Construction of
anthropomorphic phantoms for use in dosimetry studies. Journal of Applied Clinical Medical Physics [Internet]. 2009 [cited 12 April 2016];10(3). Available from: http://www.jacmp.org/index.php/jacmp/article/view/2986/1661
42. Fisher R. TISSUE EQUIVALENT PHANTOMS FOR EVALUATING IN-PLANE
TUBE CURRENT MODULATED CT DOSE AND IMAGE QUALITY [Internet]. 2016 [cited 12 April 2016]. Available from: http://etd.fcla.edu/UF/UFE0017940/fisher_r.pdf
43. Journal of the ICRU (December 2005) 5 (2): 103-113. International Commission on
Radiation Units and Measurement. Tissue substitutes in radiation dosimetry and measurement. Report No. 44. Bethesda, MD: ICRU, 1989. doi: 10.1093/jicru/ndi035
62
44. Paul A. Yushkevich, Joseph Piven, Heather Cody Hazlett, Rachel Gimpel Smith, Sean Ho, James C. Gee, and Guido Gerig. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. Neuroimage. 2006 Jul 1; 31(3):1116-28.
![Approved : Date : - [email protected] Home](https://img.pdfslide.us/doc/110x75/6207486449d709492c2fda8d/approved-date-emailprotected-home.jpg)
