SEGMENTATION AND CORRELATION OF OPTICAL COHERENCE TOMOGRAPHY AND X-RAY IMAGES FOR BREAST CANCER DIAGNOSTICS

JONATHAN G. SUN*†, STEVEN G. ADIE*, ERIC J. CHANEY* and STEPHEN A. BOPPART*†‡ *Beckman Institute for Advanced Science and Technology 405 North Mathews Avenue, Urbana, Illinois 61801, USA Department of Bioengineering Beckman Institute for Advanced Science and Technology 405 North Mathews Avenue, Urbana, Illinois 61801, USA Departments of Electrical and Computer Engineering and Internal Medicine University of Illinois at Urbana-Champaign, Urbana, Illinois, USA boppart@illinois.edu


Introduction
Breast cancer is estimated to be the leading form of new non-skin cancer cases a®ecting women in the United States in 2013. 1 As medicine evolves, new treatment options have developed, with a combination of surgery, radiation therapy, chemotherapy and hormone therapy being prescribed on an individual basis. 2 However, early detection still remains a critical factor for successful treatment and increased patient survival rates. 3,4Currently, the standard practice for screening for breast cancer includes physical examination and X-ray mammography.Suspicious lesions found on mammography are subsequently biopsied to provide specimens for pathological diagnosis, and surgical resection may follow.During surgery, resected breast masses containing tumor are frequently imaged with specimen radiography (X-rays) to help determine if the entire lesion has been resected with a su±cient margin of apparently normal tissue.Both X-ray mammography and specimen radiography rely on tissue contrast between normal and abnormal regions based on X-ray absorption and scattering.However, the introduction of other imaging modalities and tissue contrast mechanisms o®ers the potential to improve the diagnostic capabilities in the intraoperative setting.
Optical coherence tomography (OCT) was ¯rst described in 1991 5 and over the last two decades has been developed for applications in medicine, surgery, biology and materials testing. 6,7OCT can noninvasively visualize optical scattering properties arising from di®erences in refractive index within a sample area, such as from scattering particles or boundaries.OCT is rapidly becoming an established biomedical imaging technique because it uniquely o®ers noninvasive micron-level resolution capable of visualizing microscopic cell and tissue features in real time up to several millimeters deep in highlyscattering tissues.Relatively inexpensive compared to many other imaging modalities, OCT systems can be portable, and interface with a variety of beamdelivery systems such as microscopes, catheters, needles or hand-held probes.Most notably in ophthalmology, OCT has become the gold-standard for retinal imaging because of the transparent nature of the eye.Ocular disease detection has advanced with the ability to look beneath the surface of the retina, as well as visualize anterior segment structures.8À14 OCT also has applications in dentistry 15À18 and dermatology, 19À21 among many other areas.
In oncology, OCT has been used to di®erentiate between normal and abnormal breast tissue. 22,23sing OCT, it has been shown that positive tumor margins on freshly excised breast masses can be accurately identi¯ed intraoperatively during breast lumpectomy procedures. 24Furthermore, there is ongoing work involving intraoperative OCT lymph node imaging to provide real-time assessment of sentinel lymph nodes for determining tumor metastases and staging cancer.25À27 A commercial OCT system that implements interferometric synthetic aperture microscopy (ISAM) 28 has been used for in vivo imaging of the tumor cavity during breast surgeries. 29The use of other optical imaging technologies have been investigated to address the need for high-resolution assessment of tumor margins during breast cancer surgery, including elastic scattering spectroscopy, 30 Raman spectroscopy, 31 photoacoustic imaging 32 and di®use re°ectance spectroscopy, 33 among several others.While many of these have shown promise, several are hampered by factors such as point-based measurements, long acquisition times, lower spatial resolution or the generation of spectroscopic signatures, rather than image-based data.The work in this study o®ers the potential to correlate real-time intraoperative OCT ¯ndings with those obtained using pre-operative X-ray mammography and intraoperative specimen radiography.

Methods
Human breast tissue containing a portion of breast tumor (ductal carcinoma in situ) was obtained from Carle Foundation Hospital, Urbana, Illinois, under a protocol approved by the Institutional Review Boards at Carle Foundation Hospital and the University of Illinois at Urbana-Champaign.Imaging of tumor specimens was performed using each modality in succession, keeping the orientation and imaging areas correlated.Specimens were ¯rst imaged with a custom-built spectral-domain OCT system that utilized a titanium:sapphire laser (KMLabs, Inc.) as an optical source.This source provided broadband light centered at 800 nm with a 100 nm bandwidth, yielding an axial resolution of 4 m.The transverse resolution of the system was 12 m.Following OCT imaging, specimens were either maintained in the same cuvette or transferred to a second cuvette for imaging with a commercial micro-CT (X-ray) system (MicroXCT-400, Xradia, Inc.).After imaging, specimens were either placed in formalin for para±n-embedded histology or °ash frozen for frozen section histology.Four specimens (each $ 5 mm diameter) were imaged with both OCT and micro-CT, and an additional 4 specimens used for analysis were imaged with only OCT.A total of 2400 OCT images were collected and 1250 images were analyzed.A total of 2200 micro-CT images were collected and 760 were analyzed.Subsets of images were not analyzed because they were acquired at the edges of the tissue specimens and either contained a very small portion of the tissue (low signal) or were excessively noisy (high background).The determination of which raw images were used for analysis and which were discarded was based on manual inspection of each image to ensure there was su±cient signal-to-noise to visually identify the presence of tissue in the image.
In the clinical setting, digital specimen radiography systems (Faxitron Bioptics, LLC) are often used.This portable system would typically be located immediately outside an operating room where a radiologist can provide real-time feedback on a resected tissue mass.These types of systems use X-ray energies ranging from 10À35 keV.Standard X-ray mammography scans use X-ray energies from 24À32 keV, although lower energies may be used to enhance contrast at the expense of penetration depth.The micro-CT system used in this study can image at most all of the typical X-ray energies used in both X-ray mammography and specimen radiography, with an X-ray photon energy range of 20 to 80 keV.An X-ray energy of 30 keV was used to acquire the images in this study because this closely approximated the energies used in these clinical systems.The commercial micro-CT system had a machine speci¯cation resolution capability of 200 nm, but this level of detail was not apparent in the images taken with the device.Following OCT and micro-CT imaging, specimens were histologically prepared, sectioned, stained with hematoxylin and eosin and viewed under bright ¯eld microscopy.
Three MATLAB algorithms were developed to automatically segment the acquired images, modi-¯ed from community-developed algorithms from MATLAB Central. 34The ¯rst algorithm applied an amplitude ¯lter to an image based on a user-de¯ned threshold (using the im2bw function on a grayscale image).The second algorithm, a 2D spatial frequency ¯lter, applied a band-pass ¯lter in the Fourier domain (using the ®t2 and i®t2 functions) with a speci¯ed bandwidth around a center frequency, and isolated features within that range using a window function.The third algorithm applied a texture-based ¯lter, which found entropy values (using the entropy¯lt function as well as the mat2gray, im2bw and bwareopen functions) throughout a cross-sectional B-mode OCT image and made a binary decision to create a region based on similar features within a user-de¯ned correlation distance.Each ¯lter and segmentation approach was experimentally optimized by adjusting ¯lter parameters to obtain a segmentation result similar to manual visual segmentation, which served as the gold-standard for comparison.This ¯lter parameter optimization was done on a training set comprised of a total of 25 randomly selected images and a Student's t-test was performed to determine statistical signi¯cance.Then, for each segmentation approach, an additional algorithm was written to automatically read the images in a data set sequentially (using the imread function), process the images with optimum parameters and calculate an area and perimeter value for the tumor region in each image (using the regionprops function).Additionally, the texture-based segmentation algorithm had an additional optimization portion as it had a large variance in results from the initial training set.Two of the parameters (kernel size and correlation distance) were varied and the set with the lowest variance was used so that variability in results was minimized.Each of these algorithms was then applied to the remaining OCT and micro-CT image data sets (n ¼ 1225 for OCT and n ¼ 735 for micro-CT) and the results analyzed using a Student's t-test.

Manual image correlation
For accurate correlations between di®erent image sections of a specimen across modalities, the orientation and axes of imaging were maintained.The OCT scan volumes were generally smaller ($ 2 mm Â 2 mm Â 1:5 mm) compared to the micro-CT volumes ($ 5 mm Â 5 mm Â 3 mm) due to instrument scan parameters, and therefore the OCT volumes did not encompass the entire specimen.When comparing the two modalities, as well as comparing with corresponding histological ¯ndings (see Fig. 1), it can be seen that each modality serves a specialized role in imaging breast cancer.X-ray imaging prominently shows the presence of microcalci¯cations and variations in tissue density (see Fig. 1), which is not surprising, given the strong and di®erential X-ray attenuation by these features.OCT, in comparison, can provide morphological information based on optical scattering, and o®ers a di®erent type of tissue contrast that yields di®erent structural features.As seen in Fig. 1, suspicious microstructures visualized with OCT are not apparent in the micro-CT scan.However, because of the reduced ¯eld-of-view and imaging depth of OCT, applications are somewhat limited to speci¯c uses where the tissue sites have been exposed, and suspicious regions have already been identi¯ed, such as in intraoperative imaging.
3D correlations can also be useful.Commercial biomedical image visualization software (Amira, Visualization Sciences Group) was used to manually co-register di®erent morphological features.In this way, the rough orientation of the sample was obtained between modalities and features within the specimen were more easily correlated by looking at corresponding computationally extracted slices.3D volumes of micro-CT and OCT were compared on the same scale and orientation (see Fig. 1).Although the attenuation of X-rays is not signi¯cantly di®erent between ¯brous tissue and cancerous tissue, three separate tissue types were readily di®erentiable: adipose tissue, tumor tissue and microcalci¯cations.Similarly, these three tissue types were readily identi¯ed with OCT (see Fig. 1).The adipose tissue had a honeycomb network texture pattern based on the higher scattering by the adipocyte cell membranes.The tumor tissue was more homogeneous and higher scattering, and the microcalci¯cations were point-like objects that exhibited the highest amount of optical scattering.Based on these common image features between modalities, it was possible to segment and correlate images between modalities.

Optimizing segmentation methods
Following alignment of the orientation, scale and ¯elds-of-view between OCT and micro-CT data volumes, three MATLAB algorithms modi¯ed from community-developed algorithms were utilized to automatically segment images.Each algorithm made a binary decision, delineating tumor regions and areas.
The ¯rst algorithm used an intensity ¯lter to segment out tumor areas as seen in Fig. 2 (OCT and micro-CT).For the OCT data set, threshold values were optimized to 20%, based on visually determining the value that highlighted the tumor area best while minimizing the signal from the adipose tissue.For micro-CT, a threshold value of 25% was found to be optimal, as almost the entire tumor region was segmented without signi¯cant signal from the adipose tissue.
Intensity ¯ltering yielded better segmentation results for micro-CT data because of the large X-ray absorption contrast between the tumor and adipose regions.In the scattering-based OCT images, there are regions of adipose tissue that are more visible due to scattering from the adipocyte cell membranes.As such, the intensity ¯lter was not as e®ective for OCT image segmentation in this comparison.This intensity ¯lter algorithm, however, was the fastest, requiring only 0.7 s per OCT image and 0.1 s per micro-CT image on a 2.93 GHz Intel quad-core personal computer.
The second algorithm for segmentation was spatial frequency ¯ltering.This method was chosen because tumor areas are denser and have di®erent spatial frequency content in images compared to adipose tissue.This algorithm applied a band-pass ¯lter in the Fourier domain and returned a ¯ltered image.In order to optimize this algorithm, both the center frequency and bandwidth had to be optimized.Figure 3 shows this optimization for OCT and micro-CT, where arrays of images are shown, with the center frequency increasing in the columns from left-to-right, and the bandwidth increasing in the rows, from top-to-bottom.For both OCT and micro-CT images, the spatial frequency ¯lter was able to isolate the tumor area.The spatial frequencies were determined by multiplying the number of pixels by the frequency variable and dividing by the size of the image in millimeters.This segmentation algorithm required 0.95 s per OCT image and 1.8 s per micro-CT image.
The third segmentation method used a texturebased algorithm, because the entropy and texture of tumor and adipose regions should be di®erent.This approach was optimized around the variables of kernel size and correlation distance as seen in Fig. 4.This method yielded the best qualitative results for processing OCT images, because it was able to clearly isolate the tumor region.However, the parameters were highly sensitive to speci¯c images and needed to be optimized for each image, often giving erroneous results if not recalibrated for each image.Furthermore, when applied to micro-CT images, this algorithm had di±culty in distinguishing the borders of tumor regions, because the texture of the micro-CT images was universally grainy.The n1 variable identi¯ed in Fig. 4 is the width and height of the kernel size, increasing from top-to-bottom.The n2 variable is the square of the correlation distance, increasing from left-to-right.This algorithm required 2.3 s per OCT image and 2.3 s per micro-CT image.
To validate our selection of the given threshold values, we varied these threshold values slightly and re-analyzed the images in the study set to determine the e®ect that slight variations in the thresholds would have on the measured outcomes.Few differences were noted for the intensity and spatial ¯lters, validating the choice of thresholds for these ¯lters.However, more variability was noted for the texture-based ¯lter, as noted above, indicating that the use of this ¯lter approach would likely be more tissue speci¯c.

Automated segmentation
Following the optimization of these algorithm parameters based on visual inspection of segmentation outcomes compared to the original OCT and micro-CT images, automated segmentation of the larger image datasets were performed.In order to determine the e®ectiveness of each segmentation algorithm, a training set was performed with n ¼ 25 OCT and n ¼ 25 micro-CT images, comparing the area and perimeter of the segmented areas to a goldstandard of visual segmentation.Figure 5 shows the comparison of segmented areas for each method, after being normalized by the area determined by visual segmentation.Only the spatial frequency ¯lter applied to the micro-CT images was shown to be statistically similar (p > 0:05) to the area segmented by visual inspection, and it was shown to be the closest to the visual segmentation of tumor area in OCT as well.Subsequently, the spatial frequency algorithm was used in the automated portion of this study as the normalizing quantity.The intensity-based ¯lter segmented a much larger area than visual segmentation because it did not accurately account for noise in the images.The texture-based ¯lter was not able to segment a tumor area approximately 25% of the time, and also had limitations with the micro-CT images, resulting in large error bars and smaller segmented areas.This is likely due to the highly variable texture within the images, from the high heterogeneity of the tissue types.
Additionally, the perimeter of automatically segmented areas was compared to visually segmented areas (see Fig. 6).The perimeter of the tumor regions in the OCT images segmented with the intensity method was large because many small areas were isolated, each contributing to the perimeter measurement.For the spatial frequency algorithm, the perimeter was also larger than by visual segmentation because there was more detail in choosing the boundaries of the area.Finally, the Finally, based on the ¯ndings from the training set, these algorithms were applied to the remainder of the datasets, including images where tumor regions were larger and smaller.In each of these images, the intensity and texture-based algorithms were normalized to the areas found by the spatial frequency ¯lter.This was done because the spatial frequency algorithm was closest to visual segmentation in the training set (see Fig. 5).In Fig. 7, it can be seen that intensity-based segmentation consistently found larger areas than those identi¯ed by the spatial frequency algorithm, as shown earlier.
However, the texture-based method performed better for OCT than micro-CT.This is consistent with previous ¯ndings, but the texture-based method has the additional advantage of ¯nding smaller numbers of isolated areas.
In the future, algorithms can be developed to automatically determine and evaluate threshold settings based on speci¯c breast cancer sub-types and image features.Also, a combination of segmentation and ¯lter methods may be useful for automatically identifying areas of tumor so that real-time segmentation of OCT can be used in the operating room without the need for operator training to read and interpret the OCT images.Additionally, these algorithms can be expanded to segment 3D datasets to isolate 3D volumes of tumors.

Conclusions
Intraoperative OCT has been shown to identify positive tumor margins during breast cancer surgery.Because X-ray mammography and specimen radiography are typically used in the screening, diagnosis and treatment of breast cancer, this study sought to segment, co-register and correlate imagebased ¯ndings of breast cancer observed with OCT and X-ray imaging (micro-CT).This multi-modal (and multi-scale) imaging approach has the bene¯t of leveraging the large volume of image data acquired using X-ray imaging, and relating this information to 0 0.5 the high-resolution real-time OCT.Future work will extend the ¯ndings of this study to intraoperative OCT, from either resected tumor specimens or even in vivo within the tumor cavity.This study compared the e®ectiveness of three algorithms for automatically segmenting OCT and micro-CT images of human breast tissue and tumors.These algorithms were based on intensity, spatial frequency and texture-based segmentation approaches.The spatial frequency algorithm had the most consistent segmentation of tumor areas.However, the intensity and texture-based algorithms each had advantages that could be incorporated into a more sophisticated segmentation algorithm in the future.Quantitatively determining the ability to segment and co-register images between OCT and X-ray imaging, and automating the registration process, will o®er the potential to more comprehensively evaluate tissue for diagnostic purposes.

Fig. 1 .
Fig. 1.Representative comparison of OCT, micro-CT and histology image data.The dotted white boxes in the micro-CT images and the dotted black boxes in the histology images (H&E) are the approximate scan areas of the corresponding OCT images.Tumor structures can be seen in all three modalities on the left-half of the images, with normal adipose tissue on the right-half.The arrows indicate corresponding structures and areas of microcalci¯cations.3D OCT and micro-CT datasets are also shown.The black 3D wire-frame in the micro-CT image is the approximate scan area for the corresponding 3D OCT image.Microcalci¯cations (black arrows) appear bright in the micro-CT data of the adipose tissue on the right.The micro-CT volume is visualized using a direct volume rendering algorithm while the OCT volume is visualized using a surface mesh algorithm.

Fig. 2 .
Fig. 2. Optimization of intensity-based ¯ltering.OCT images (top set) with intensity threshold settings from 5À30% and micro-CT images (bottom set) with threshold values from 15À40% are shown.Setting threshold values at 20% for OCT and 25% of maximum intensity for micro-CT yielded the best results by visual selection (boxes).The image dimensions were 3 mm Â 1:5 mm and the tumor area is shown on the left side of each image.

Fig. 3 .
Fig.3.Optimization of spatial frequency ¯ltering.In both sets of OCT and micro-CT images, the tumor tissue is located on the left.The center spatial frequencies vary in the range of 73.4À75.8cycles/mm from left-to-right for OCT and 18.4À32.2cycles/mm from left-to-right for micro-CT.The bandwidth varies in the range 1.4À2.6 cycles/mm from top-to-bottom for OCT and 6.3À15.7 cycles/mm from top-to-bottom for micro-CT.The optimal set of parameters has a center spatial frequency of 74.9 cycles/mm and a bandwidth of 2.2 cycles/mm for OCT, and a center spatial frequency of 23.01 cycles/mm and a bandwidth of 12.3 cycles/mm for micro-CT (boxes).

Fig. 4 .
Fig. 4. Optimization of texture-based ¯ltering.In both sets of OCT and micro-CT images, the tumor tissue is located on the left.The columns show increasing correlation distance (n2) from left-to-right and the rows show increasing kernel size (n1) from top-tobottom.The optimum con¯gurations are selected visually (boxes).
Fig. 5. Average area of tumor found by each segmentation method, normalized to visual segmentation (training dataset, n ¼ 25) with error bars representing standard deviation.A Student's t-test comparing the mean of areas found by each method with the mean obtained by visual inspection was performed.The p-values are shown in the table.Note, because the data is normalized, areas closer to 1 more closely approach the area determined by visual segmentation.Average perimeter of tumor found by each method normalized to visual segmentation (training dataset, n ¼ 25) with error bars representing standard deviation.A Student's t-test comparing the mean of perimeters found by each method with the perimeter obtained by visual inspection was performed.The p-values are shown in the table.Note, because the data is normalized, perimeters closer to 1 more closely approach the perimeter determined by visual segmentation.Comparison of automatic segmentation algorithms.Areas of tumor segmented by each algorithm are normalized to areas from the spatial frequency algorithm (n ¼ 1250 for OCT, n ¼ 760 for micro-CT), with error bars representing standard deviation.A Student's t-test comparing the means of the areas found by each method with the area obtained by the spatial frequency ¯lter was performed.The p-values are shown in the table.