Terahertz Imaging of Three-Dimensional Dehydrated Breast Cancer Tumors
- 491 Downloads
This work presents the application of terahertz imaging to three-dimensional formalin-fixed, paraffin-embedded human breast cancer tumors. The results demonstrate the capability of terahertz for in-depth scanning to produce cross section images without the need to slice the tumor. Samples of tumors excised from women diagnosed with infiltrating ductal carcinoma and lobular carcinoma are investigated using a pulsed terahertz time domain imaging system. A time of flight estimation is used to obtain vertical and horizontal cross section images of tumor tissues embedded in paraffin block. Strong agreement is shown comparing the terahertz images obtained by electronically scanning the tumor in-depth in comparison with histopathology images. The detection of cancer tissue inside the block is found to be accurate to depths over 1 mm. Image processing techniques are applied to provide improved contrast and automation of the obtained terahertz images. In particular, unsharp masking and edge detection methods are found to be most effective for three-dimensional block imaging.
KeywordsBiomedical optics Medical imaging Breast cancer Terahertz imaging
Breast conservation surgery, also called lumpectomy, involves the excision of a breast cancer tumor with a margin of healthy tissue. The excised tumor is then processed by a pathologist, which could take several days, in order to determine whether there is any cancer remaining on the surgical edge, denoting a positive margin . Once positive margins are detected, a second surgery is required to remove the remaining cancerous tissues. Even with modern techniques, positive margin rates are reported to be as high as 20–40% . To minimize the need for second surgery, it is necessary to develop a rapid and accurate intraoperative method for margin assessment . Terahertz (THz) imaging is proposed here to investigate the margins of excised tumors.
While THz imaging has proven potential providing contrast between breast cancer and healthy tissue in both fresh and formalin-fixed, paraffin-embedded (FFPE) tumors [4, 5, 6, 7], all published work was performed on flat sections of the tumor. While the ultimate goal is to investigate the margins of freshly excised tumors, here, we focus on excised dehydrated tumors fixed in formalin and embedded in paraffin. To the authors’ knowledge, this is the first investigation for in-depth imaging of breast tumors using THz to approach the problem of margin assessment. In addition, the investigation of paraffin tissue blocks could provide useful information for pathologists to determine the position and extent of the embedded tissue prior to the histopathology sectioning. Some of our preliminary results were presented in conference papers [8, 9]. Ongoing research is investigating THz imaging of freshly excised tumors grown in mice.
THz imaging technology has been a rapidly expanding field of research for biomedical applications in recent years . THz imaging and spectroscopy applications have expanded rapidly with the development of THz sources and systems , and research in the 0.1 to 4 THz range has been applied to a wide variety of medical conditions, showing clear contrast in assessment of liver cirrhosis , myocardial infarction , burn wounds , and cancer diagnosis . THz is an attractive approach for biomedical applications due to having higher resolution than microwave frequencies while being shown to penetrate over a millimeter in fat  and through several millimeters in fixed tissue . Additionally, THz imaging is sensitive to water content in tissue  and uses non-ionizing radiation such that it is biologically safe for in vivo applications . THz imaging has been successfully applied to cancer of the liver [13, 17], colon , brain , skin , and breast [4, 5]. While the THz sensitivity to water content is one source of contrast between different kinds of cancer and adjacent healthy tissue, several investigations have shown clear differentiation of dehydrated tissues as well, showing the strong potential of THz in assessment of cancer [4, 17, 21].
In addition, image processing techniques are implemented to enhance the THz images. Within an intraoperative setting, the use of automated image generation and signal-based cancer detection will reduce both staff training needed to use the THz imager and observer bias in determining the status of the margins. Basic image processing techniques can greatly improve the visualization of THz images using intensity windowing and histogram manipulation [23, 24]. For example, the use of edge detection and region growing techniques are implemented to segment images into regions of cancer and healthy tissue [23, 25]. One automated approach for distinguishing breast cancer tissue from normal tissue region makes use of data reduction techniques together with support vector machine (SVM) classification with radial basis functions to distinguish tumor from normal tissue in excised breast tissue samples . Several methods for data reduction were explored. The first utilized ten heuristic parameters that characterized time domain and frequency domain properties of the THz signals, the second made use of principal component analysis (PCA) of the THz pulses, and the third made use of the PCA of the ten heuristic parameters. The authors found that SVM classification of the top ten principal components yielded 92% tissue classification accuracy for the 51 tissue samples in their study. This is a very strong result given the amount of data reduction performed (from 512 time samples of the THz signal down to 10 PCA coefficients) . Our group has performed preliminary work on image processing applied to THz images of breast cancer tissue to aid in the visibility of important image features such as the boundaries of cancer tissue and normal tissue . In this work, we explore additional image enhancement, edge detection, and image segmentation methods for THz images to aid in automated tissue classification.
The layout of this work will be as follows: Section 2 will address tissue sample preparation, and a description of the THz system; Section 3 will discuss signal processing used in THz images; Section 4 will present results of THz images and image processing; and Section 5 will include concluding remarks and future investigations.
2 Tissue Sample Preparation and THz Imaging System
The breast cancer tissue samples used in this work were purchased from the biobank at the National Disease Research Interchange (NDRI) or obtained from Northwest Arkansas (NWA) Pathology Associates, P.A. Samples were obtained as three-dimensional (3D) bulk FFPE tissue embedded in paraffin blocks. Following THz scanning of the tissue in blocks, the samples were sectioned at NWA Pathology into 30-μm thick slices and were mounted on glass or polystyrene slides. Histopathology assessment was performed on 5-μm thick sections sliced between each thicker sections of 30 μm and stained with hematoxylin and eosin (H&E) for validation of THz images. The tissue sections used here will be classified as follows: sample 1 was obtained from a 54-year-old patient diagnosed with grade III/III infiltrating ductal carcinoma (IDC), sample 2 was obtained from a 39-year-old patient diagnosed with grade III/III IDC, and sample 3 was obtained from a 69-year-old patient diagnosed with grade II/III lobular carcinoma (LC). Samples 1 and 2 were provided by NDRI and sample 3 was provided by NWA Pathology Associates. Histopathology images of the samples will be compared with THz images.
3 Signal and Image Processing
3.1 Time of Flight Signal Analysis
where t 0 is the location of the primary peak of the reflected signal at each pixel.
The estimation of (4) relies on measuring a secondary peak and hence the time distance∆t as shown in Fig. 2. In the event that one of the tumor regions (i.e., carcinoma, fatty, or fibroglandular) is very close to the paraffin surface, the secondary signal from that particular region could be merged into the primary peak, which makes estimating the depth of that particular region not possible. Also, as will be presented in the results of Section 4, the secondary reflection will not be seen at pixels located inside the same tissue region inside the tumor; however, once the tissue region changes, the secondary reflection can be seen in the measurements indicating to a change in the tissue type.
3.2 Image Processing Techniques
The image processing in this work converts time domain THz signals from the measurement system’s file format (TeraView’s TVL) to three-dimensional RAW image files so they can be visualized using the open source software package MeVisLab . Since the TPS Spectra 3000 THz system uses step motors without encoders when collecting time domain signals, there is some horizontal alignment error between even and odd rows of the obtained images. To remove these alignment errors, we shifted odd rows in the image by minimizing the mean squared error (MSE) between the time domain signals of each odd row and its two adjacent even rows. This aligned image is then used for further image processing.
3.2.2 Intensity Mapping
The result of this transformation is that intensity values with fewer points in the image (less relevance) are grouped together and intensity values with more points in the image (greater relevance) are distributed across the intensity range, providing greater contrast when the difference between tissue regions is relatively small.
One other method used in this work for visualization of the three-dimensional scans is uniform scaling. Rather than perform a histogram equalization, the intensity of the x-y cross section at each time domain point is considered individually and rescaled from 0 to 1 based on the local maximum and minimum. Since secondary reflections inside of the tissue block are likely to be reduced from the transmission and reflection losses of the signal, this gives equal consideration to primary surface reflections and later secondary reflections in the visualization of the scan.
3.2.3 Edge Sharpening
Another method for improving the visualization in THz imaging is the use of edge sharpening . This is done by adding a secondary mask to the original image for all three dimensions of the scan such that output(x, y, t) = input(x, y, t) + α × mask(x, y, t), where α is an adjustable control variable. The mask itself is obtained by subtracting a blurred image from the input, mask(x, y, t) = input(x, y, t) − blurred(x, y, t), where the blurred image is obtained by performing neighborhood averaging of the N × N × N points centered around the position being solved. Gaussian averaging can also be used for this purpose but was not found to show any significant difference from neighborhood averaging in this work. Edge sharpening using the blurred image to create the mask is referred to as unsharp masking.
3.2.4 Edge Detection
Another robust method for performing edge detection is a classic technique known as Canny edge detection . This technique works by smoothing the input image using a Gaussian filter and then finding the zero crossings of the second derivative along the gradient direction, which correspond to the maxima/minima of the first derivative. Non-maxima suppression is then used to remove zero crossings corresponding to minima of the first derivative. Next, the gradient magnitude of the image is calculated using the Sobel operator. Any gradient value greater than an assigned threshold denotes a strong edge, while any points adjacent to a strong edge that meet a lower threshold are considered a weak edge. Finally, connectivity analysis is used to connect any strong and weak edges found in the same 3 × 3 neighborhood of points. Existing Canny edge detection algorithms were implemented to obtain the images in this work .
3.2.5 Region Growing
Edge detection is useful for visualizing the outlines of objects in images, but in some cases there are weak or broken edges in an image, and these outlines do not fully enclose objects of interest in an image. In these cases, image segmentation techniques that focus on pixel similarity are often more effective. One classic segmentation algorithm is region growing, which starts with one or more seed points in the image defined to be cancer or healthy tissue. Then adjacent points in the image are compared to the seed points and added to the region for the defined tissue if they are found to be similar enough to the original points based on a predefined threshold or other criteria . For our THz images, we have 1024 time domain samples at each (x,y) location. There are three natural choices for comparing these sample vectors: to calculate their correlation (inner product), to calculate the L1 norm (sum of absolute differences), or to calculate the L2 norm (sum of square differences). A fourth option is to compare the intensities of the peak value in each THz signal. In our experiments, the L2 norm was found to be most successful in growing regions corresponding to cancer or healthy tissue.
4.1 Image Processing Results
4.1.1 Sample Preparation
The above discussed image processing techniques are first tested on tissue sections of 30-μm thickness taken from sample 1 and sample 2 and mounted on microscope slides. The section from sample 1 is mounted on glass, while the section for sample 2 is mounted on polystyrene. THz x-y reflection imaging is taken for both samples, and the resulting images are used in testing the improvement of image processing techniques. The H&E stained slides are also taken adjacent to these 30-μm sections to obtain histopathology images for validation. For the imaging performed in this work, the original THz scan of the samples took approximately 35 min at a step size of 200 μm. All additional processing of the dataset took less than a minute.
4.1.2 Intensity Mapping
4.1.3 Edge Sharpening
4.1.4 Edge Detection
4.1.5 Region Growing
The histopathology image of sample 2 is shown in Fig. 8d. Here, the region grown from the intensity peak in Fig. 8e as well as the region obtained from L2 mapping in Fig. 8f both show very good detection of the cancer region in these areas. These grown regions even account for the necrosis region at the center of the IDC and show a high level of accuracy against the histopathology. Therefore, while some edge detection techniques were unable to reliably resolve the boundaries of the tissue regions, a region growth method has proven effective at obtaining accurate regions that can be ascribed to the cancer in the tissue.
4.2 Three-Dimensional Imaging of Breast Cancer Tissue Blocks
4.2.1 Sample 1: Infiltrating Ductal Carcinoma
Figure 10b shows reflected signal from the surface of the faced off block with the tissue exposed in order to correlate the tissue regions to the histopathology. Here, the tissue regions are clearly defined, with the IDC on the right and the fibroglandular on the left. In the faced off block, the two regions are not distinct at the surface, but the outside border of the tissue is clear. From the THz scan with the paraffin covering the tissue, the reflected peak from the surface of the tissue inside the block is shown in Fig. 10c. This image is obtained using the values of the secondary peak of the signal at each pixel in the image as discussed in Section 2. Here, a significant reflection can be seen over the area of IDC on the right consistent with the histopathology image in Fig. 10a, though the full extent of the region is not as clearly outlined as in Fig. 10b. Additionally, the carcinoma shows clear contrast compared to the region of fibroglandular tissue on the left, which shows smaller reflection values. This layer can be more clearly visualized by making use of cross section images into the depth of the block by observing the x-z or y-z planes of the scan, with the z-axis corresponding to the time domain of the measured reflection signals using the time of flight estimation technique in (5) to provide an approximation of feature depth beneath the surface of the block. This scan with the estimated z-axis depth will be referred to as the Z-scan. The dashed lines intersecting at point A in Fig. 10c show the positions where the cross section images are taken, and further clarification is given in Fig. 10d, which shows a 3D diagram of the scan (i.e., x-z and y-z planes). Since the THz Z-scan produces a 3D dataset, these additional cross section images take no additional time to acquire. The x-z cross section image is shown in Fig. 10e, and the y-z cross section image is shown in Fig. 10f. The figures are oriented in the same direction as the experimental setup, with the signal coming from below to reflect off the sample such that the tissue is above the air-paraffin interface. These images make use of uniform scaling at each depth in order to highlight the secondary reflections. However, the uniform scaling increases the noise in the air in Fig. 10e–j due to the very low signal in this region, and there is some degree of noise in the paraffin block away from the tissue interfaces as well. Reflections arise when an interface between tissue regions or between tissue and the paraffin block is encountered by the THz signal moving in the z-direction. In both cross section views, the reflection from the block surface, from the top of the tissue, and from the bottom of the tissue are all clearly visible for the tumor, while the side walls of the tissue are at a more oblique angle with the THz signal and do not appear. The interface between paraffin and the top of the tumor is estimated using (4) to be between 150 and 200 μm, while the bottom of the tissue has a range between 1 and 1.5 mm. In contrast, the fibroglandular tissue region shows some distributed scattering but no clearly defined reflections outlining the entire region. This is likely due to a high similarity of the dehydrated fibroglandular tissue to the surrounding paraffin.
In order to investigate enhancement and automation of the THz imaging processing, unsharp masking and edge detection are applied to the THz scan due to their effectiveness in the imaging of tissue sections. It should be noted that while region growing is shown to be effective for tissue sections in Fig. 8, it is not found to resolve the three-dimensional block imaging well and requires more work to be implemented. The results of the unsharp mask method can be seen for the x-z and y-z cross sections in Fig. 10g, h, respectively. Here, the reflections from the tissue top and bottom are defined more clearly, and many of the horizontal effects in the block not corresponding to the tissue reflections are diminished. Thus, the unsharp mask shows good clarification of the tissue boundaries while decreasing other effects in the signal. This effect can be seen more clearly using the automated Sobel operator as seen for the x-z and y-z cross sections in Fig. 10i, j. It can be seen that any signal in the block aside from the tissue reflections is suppressed, leaving the clear reflections from the top and bottom of the tissue. This technique also highlights the scattered reflections through the depth of the fibroglandular tissue, as seen on the left side of Fig. 10i. As a result, the boundaries and margin of the infiltrating ductal carcinoma are clearly determined in the 3D THz scan of the paraffin block in Fig. 10, and image processing shows good results in improving the visibility of the tissue at depth.
The unsharp mask and edge detection processing techniques are implemented on the THz data as shown in Fig. 11 in the third and fourth columns. The unsharp mask image in Fig. 11c shows sharper details with better contrast in the tissue reflections compared with Fig. 11b. The edge detection using the Sobel operator in Fig. 11d clearly outlines the region of IDC at the surface of the tissue and provides edges of the more scattered fibroglandular tissue. Since there is no significant reflection at z = 180 μm in the THz image in Fig. 11f, there is likewise no significant effect of the image processing in Fig. 11g, h. As the bottom reflection becomes visible in Fig. 11j, the unsharp mask method in Fig. 11k shows some improvement in the details of the bottom reflection, while the edge detection in Fig. 11l shows excellent definition of the tumor edge as the dark red line. The image processing shows similar improvement at z = 850 μm, where the unsharp mask results in Fig. 11o show improved feature resolution over the standard THz image in Fig. 11n, and the edge detection in Fig. 11p accurately defines the bottom edge of the receding tumor.
The results of Fig. 11 show the effectiveness of THz in detecting the boundaries of cancerous tissues buried in the paraffin block. These 3D THz images provide insight into the interaction of the THz signal with the heterogeneous tumor tissues.
4.2.2 Sample 3: Lobular Carcinoma
The surface reflection from the tissue in the paraffin block is given in Fig. 12b, where the lobular carcinoma shows a distinctly higher reflection from the rest of the tissue, with the fibroglandular showing slightly lower reflection and the more fatty tissue appearing only slightly different from the surrounding paraffin block. The dashed lines intersecting at point A indicate the cross sections selected for looking at the tissue in-depth, which is further clarified in the 3D diagram in Fig. 12c. The in-depth cross sections of the block can be seen in Fig. 12d for the x-z view and Fig. 12e for the y-z view of the dashed lines in Fig. 12b and imaging planes in Fig. 12c. Since the block was faced off prior to scanning, the cancer tissue is already present, the tissue surface reflection is aligned with the block reflection. The reflection from the bottom of the tissue is estimated to be between 1.5 and 2 mm, though part of the reflection is seen to extend beyond the range of the Z-scan. The reflection from the bottom of the tissue is broader along the z-axis than the reflections in sample 1 due to the increased depth of the signal in the paraffin block. The use of unsharp mask enhancement in Fig. 12f, g shows some resolution improvement of the tissue reflection with decreased horizontal smearing but slightly increased noise. Similarly, the automated edge detection in Fig. 12h, i shows clear definitions of the boundary, including the very slowly receding edge of the LC in Fig. 12i. In all cases, the boundary of the cancer tissue is clearly defined at depth, with image processing showing greater clarity.
The unsharp masking and edge detection methods applied to the THz data of sample 3 are shown in the third and fourth columns of Fig. 13, respectively. Unsharp masking in Fig. 13c shows enhanced features compared to the THz image in Fig. 13b, with sharper edges and even some definition for the fibrous streaks in the fibro/fatty tissue region. Edge detection using the Sobel operator is shown to clearly define the outline of the tissue in Fig. 13d, as well as defining the border between lobular carcinoma and the fibro/fatty tissue. Since there are only small reflections at z = 1050 μm in Fig. 13f, the unsharp mask in Fig. 13g shows relatively little improvement in the imaging. Edge detection in Fig. 13h shows clear definition of the edge of the fibro/fatty region along with some faint resolution of the other tissue boundaries. At z = 1620 μm, the use of unsharp masking in Fig. 13k shows slight improvements in the visualization of the reflection in Fig. 13j but also increases the noise surrounding the tissue. The edge detection in Fig. 13l clearly distinguishes the reflection from the bottom of the tissue. It should be noted that there are some additional reflections from the edge of the fibro/fatty region with the paraffin block that continue to arise in these images and are clarified by the automated edge detection. As the depth increases to z = 2000 μm, slight improvement of the reflection edge can be seen using the unsharp mask in Fig. 13o, though due to the scattered nature of the reflection the increased noise in the surrounding block becomes a problem in resolving the reflection. Likewise, Fig. 13p shows noise around the final reflection but clearly resolves the tissue bottom.
The results in Figs. 12 and 13 show the effectiveness of THz for imaging lobular carcinoma and demonstrate a penetration depth of at least 2 mm. Thus, the potential of THz imaging for various pathologies of breast cancer is clearly demonstrated.
This work showed the successful application of THz imaging to both infiltrating ductal carcinoma and lobular carcinoma embedded in paraffin blocks. THz imaging showed clear definition of the upper and lower boundaries of cancer in the block, which was correlated in 3D with histopathology sections sliced throughout the blocks. While the histopathology images showed the tumors through at any section throughout the block, THz imaging highlighted the boundaries of the cancer only when a change in tissue type occurred. Furthermore, the 3D imaging of the blocks could be segmented into x-y, x-z, and y-z cross section images in order to visualize these boundaries electronically without the need for slicing the tissue. These results show the effectiveness of THz imaging for the assessment of tumor margins, where cancer tissue is near the edge of the surgical excision.
Image processing techniques were shown to be effective for THz images of tissue sections and three-dimensional tissue embedded in paraffin blocks. Several methods showed image improvement mostly for flat sections of breast cancer tissue. However, unsharp masking and edge detection techniques were shown to be effective for the images of the three-dimensional tissue in blocks. In particular, edge detection using a Sobel operator showed very good definition of the cancer boundaries. The overall enhancement provided by these techniques is not significant, indicating that the manual methods were successful at the expense of training and time by a system operator. The image processing techniques critically provide automation for THz imaging without the need for training the operator. These techniques lend the THz imaging to be used within an intraoperative setting.
The authors would like to thank the pathology staff at the Northwest Arkansas Pathology Associates, P.A., for providing histopathology services for the tissue used in this work. This work was funded by NSF-MRI no. 1228958 and NSF awards no. 1408007 and no. DGE-1450079, and the University of Arkansas Distinguished Doctoral Fellowship program.
- 1.L. Jacobs, Ann. Surg. Oncol. 15, 5 (2008).Google Scholar
- 6.T.C. Bowman, Experimental Terahertz Imaging and Spectroscopy for Ex-vivo Breast Cancer Tissue, University of Arkansas (2014).Google Scholar
- 8.T. Bowman, M. El-Shenawee, L.K. Campbell, Proc. SPIE 9706, (2016) doi: 10.1117/12.2211167.
- 10.S. Fan, Y. He, B.S. Ung, E. Pickwell-MacPherson, J. Phys. D Appl. Phys. 47, 37 (2014).Google Scholar
- 13.Y. Miura, A. Kamataki, M. Uzuki, T. Sasaki, J. Nishizawa, T. Sawai, Tohoku J. Exp. Med. 223 (2011).Google Scholar
- 15.C. Yu, S. Fan, Y. Sun, E. Pickwell-Macpherson, Quant. Imaging Med. Surg. 2, 1 (2012).Google Scholar
- 20.P. Doradla, K. Alavi, C. Joseph, R. Giles, J. Biomed. Opt. 18, 9 (2014).Google Scholar
- 21.S.J. Oh, S.-H. Kim, Y.B. Ji, K. Jeong, Y. Park, J. Yang, D.W. Park, S.K. Noh, S.-G. Kang, Y.-M. Huh, J.-H. Son, J.-S. Suh, Biomed. Opt. Express 5, 8 (2014).Google Scholar
- 23.R.C. Gonzalez, R.E. Woods, Digital Image Processing, 3rd edn. (Pearson Prentice Hall, NJ, 2008), pp. 120–144, 689–794.Google Scholar
- 25.J.F. Canny, IEEE T. Pattern Anal. 8, 6 (1986).Google Scholar
- 26.Y. Wu, T. Bowman, J. Gauch, M. El-Shenawee, Proc. SPIE 9706 (2016) doi: 10.1117/12.2209706.
- 27.MeVisLab official website. (2015). Available: http://www.mevislab.de/
- 28.Detect Edges with Canny Edge Detection Filter (ITK, 2015) http://itk.org/ITKExamples/src/Filtering/ImageFeature/DetectEdgesWithCannyFilter/Documentation.html (accessed 26 August 2016).
- 30.M. Naftaly, Terahertz Metrology, 1st edn. (Artech House, MA, 2015).Google Scholar