AUTOMATED FEATURE ANALYSIS IN BIOLOGICAL IMAGES

By

Guneet Singh Mehta

A thesis submitted in partial fulfillment of

the requirements for the degree of

Master of Science

(Electrical and Electronics Engineering)

at the

UNIVERSITY of WISCONSIN-MADISON

2017

Abstract This thesis is comprised of three projects that I worked on during the period of my Master of Science degree. All three projects use computer vision and image processing techniques to improve microscopy image analysis workflows and develop object detection applications. This work also discusses an image analysis tool developed for imaging and analysis of collagen. The first project is aimed at improving the current state of image acquisition by autofocusing the slide and removing artifacts from image by flat field correction. This project will serve as a stepping stone for smart microscopes where runtime analysis can be done during acquisition. The second project was developed in collaboration with the Exploratorium Museum (San Francisco) to detect and highlight zebra fish embryos and zebrafish in a stream of video captured by a microscope as the objective is moved or zoomed by users. The aim of this project was to improve museum visitor participation by highlighting all the zebrafish in current field of view. The third project is aimed at developing data analysis and data visualization tools which use the fiber data extracted from Second Harmonic Generation (SHG) images by CT-FIRE (Curvelet Transform - Fiber Extraction Algorithm) software. Two broad functionalities developed were: Post Processing Graphical User Interface (GUI) for fiber analysis; and Region of Interest (ROI) manager.

Introduction

This work is based on multiple projects with a common theme of finding meaningful features in biological images and converting them into numbers for inference and hypothesis testing. The ground for this work was laid in 2014 during my internship at Laboratory for Optical and Computational Instrumentation (LOCI) at University of Wisconsin-Madison under LOCI Director Dr. Kevin Eliceiri and Co-Advisor Prof Dan Negrut. My summer experience gave me a broad overview of imaging techniques, computer vision, image processing and software developed at LOCI. My summer work also exposed me to ImageJ software, acquisition tools and MATLAB tools for data analysis. I also gained hands-on experience in developing and maintaining the MATLAB based suite of software for fiber analysis in Second Harmonic Generating (SHG) images. Subsequently, I decided to join LOCI as a Research Assistant while pursuing my Master of Science degree at Electrical and Computer Engineering Department at UW Madison. While trying to solve the research problems, I got to learn about ImageJ software, ImageJ Macros, building standalone MATLAB projects and using libraries for software development. This thesis resulted in two published papers titled : “Comparison of Picrosirius Red Staining With Second Harmonic Generation Imaging for the Quantification of Clinically Relevant Collagen Fiber Features in Histopathology Samples” and a book chapter “Liu, Y., Keikhosravi, A., Mehta, G.S., Drifka, C.R., and Eliceiri, K.W. (2017). Methods for quantifying fibrillar collagen alignment. In Fibrosis: Methods and Protocols, L. Rittié, ed. (New York: Springer) (in press)” This work includes details of the projects I have tackled during my degree, which are organized in three chapters, concluding remarks and appendices for further reference.

Chapter 1 Smart Microscope

One of the common workflows of biologists while imaging is to image a lot of samples in a similar manner. Current systems require human intervention to change slides, reset the objective, post process acquired images etc. Automation of these routine tasks will remove humans from the decision making loop and improve the speed of acquisition. One of the software systems at Laboratory for Optical and Computational Instrumentation (LOCI) University of Wisconsin-Madison, is called WiscScan which is an open source project for microscope control and imaging. As initial steps to reduce human intervention and post acquisition image processing, the following two functionalities were implemented in C++:-

1. Autofocus during image acquisition 2. Flat field correction during image acquisition

Part A: Autofocus during image acquisition Current imaging systems while imaging large tissues, acquire multiple images at the same Z-stack level. These individual images (also referred as tiles) are then stitched together to create a bigger image. Many a times the tissue being imaged has different thickness at different points (due to improper slicing), requiring different focus at different portions of the tissue. In current systems, at the start of imaging, the scope is focussed such that the first tile is under focus and during acquisition the stage is moved only in x and y directions (in plane of slide) and not in z direction (perpendicular to slide plane). Thus the focus setting of the first tile is used for all other tiles, which can often lead to acquisition of a blurred image. This work proposes to use a common numerical measure of focus called Focus Measure to correct the Z-position of the slide as it is moved in raster scan order while acquisition. The Focus Measure is a measure of high frequency content / Entropy in the image. An out of focus image (Figure 1) is blurry and does not have sharp edges, while a focussed image will have sharp edges(Figure 2). Focus Measure for an image is calculated as:-

Focus Measure=

(A) (B)

Figure 1: Images with different focus settings: (A) Out of focus image and (B) In focus image where

And (i,j) is the position of each pixel. The second derivative in x and y directions is calculated as :-

Focus measure for images at different Z positions are calculated and the z position with maximum Focus Measure is the best focussed image. For example the following images are acquired at increasing z distance of the stage from the objective (in raster scan order)

A B

C D

E F

G

Figure 2:- Images with increasing z distance : A,B,C,D,E,F,G

Focus Measure of images

Image Focus Measure

A 29.3561

B 64.6141

C 82.5003

D 67.2039

E 39.7218

F 19.406

G 10.3512

Table 1: Focus measure of images in Figure 2

It can be seen from the above images that image “C” has maximum focus measure. It can also be visually verified that image “C” is the most focussed image with sharpest edges and features. The code for the above functionality is attached in Appendix A Part B: Flat Field Correction during image acquisition Flat Field correction is a standard calibration technique which is used to remove the effects of system defects and errors from a digital image. When a perfectly uniform background is imaged using a microscope, different regions in the acquired image may not have the same pixel values due to optical path distortion or different sensitivities of individual pixels in a camera pixel array. Once an imaging system has been flat field corrected a uniform background will create an image with the same value for all pixels in the image. Thus any variation in the image acquired after correction will be a result of the sample being imaged and not any system error. For example, consider the following two images - a) Image captured without flat field correction and b) image captured with a uniform background

(A) (B)

Figure 3: Images captured without flat field correction: (A) image of a section of an H&E stained slide and (B) image of a uniform background

It can be seen that due to dust particles and non uniformity in optical paths the background image (Figure 3B) is not uniform. Image captured without compensating for this effect contains artifacts. It can be seen in the image below that such an image has dust particles and non gray border at the edges:-

Figure 4: Image containing circular brown artifacts (marked in circles) and non grey

background (marked in rectangle)

Artifacts like these lead to suboptimal performance of other algorithms like automated object detection. For example, the image below was stitched together without doing the flat field correction of individual images, leading to artifacts of horizontal and vertical stripes in the image below.

Figure 5: Image obtained after stitching images not compensated by flat field correctness.

This effect can be removed by flat field correction which requires the following:-

1. A Background image captured by imaging a perfectly flat and featureless area in a slide prior to any imaging.

2. The background image is converted to a grayscale image and average value of all pixels is calculated, referred as “mean_background_value”

3. Each pixel in each channel is traversed and the value is corrected by the formula:-

new_image[i][j][c]=raw_image[i][j][c]/background_image[i][j][c]*mean_background_value where:-

a. “i,j,c” are the pixel coordinates along x y and z axis. b. “raw_image” is the image of sample without any correction c. “background_image” is the image of a uniform featureless region in a slide d. “new_image” is the flat field corrected image. e. Division is a long form division

Results Below is the original raw image with artifacts marked in red and the flat field corrected image with the C++ code attached in Appendix B.

(A)

(B)

Figure 6: Images before and after flat field correction:(A) Raw image and (B) Flat field corrected image

Chapter 2 Detecting Zebrafish in video frames

This work led to a collaboration with the Exploratorium Museum at San Francisco to develop a computer vision system to detect zebrafish embryos and adults in a live video feed. The video was captured by a microscope mounted over a slide containing zebrafish embryo in various stages of development as shown in Figure 8.

Figure 7: Example Images of zebra fish in various stages of development

Two detectors were created for this project :- Embryo Fish Detector and Adult Fish Detector Embryo Detector The distinguishing feature of all embryo fish is its spherical shell as seen in Figure 8 below. Thus detecting circles in the image will detect position of embryos in the image.

(A) (B)

Figure 8: Embryo fish in different stages of development: A Semi developed embryo and (B) Embryo in nascent stage

However some images also contain spherical artifacts like air bubbles at the boundary of the image and empty Embryo shells as seen in the figure (Figure 9) below.

(A) (B)

Figure 9: Spherical artifacts in acquire images: (A) Air bubble at edge of the slide and (B) Empty zebra fish shell

Method: As mentioned earlier, all the embryos in the acquired images were highly circular with a specific range of radius, thus reducing the problem of finding embryos to finding circles of certain radius range (35-45 pixels in steps of 2 pixels at a Magnification of 1X in image feed of size 200 x 129 pixels) in the image. A common technique of finding regular shapes in images is using a Hough Transform. An image containing a circle is converted to an image with same size with intensity values corresponding to the probability of presence of a center of circle at that position. For example consider the following images:-

(A) (B)

Figure 10: Image and Hough Transform: (A) Image containing two circles (B) Hough Transform of (A) with intensity proportional to the possibility of presence of a circle

In practice each image (Figure 11A) is converted into an edge image containing only object outline (Figure 11B).Thereafter the edge image was converted into Hough Transform where intensity of each point is indicative of probability of presence of a circle centered at that position(Figure 11C). Using the Hough Transformed image, at most 10 most prominent circles are selected based on intensity of Hough Transform. The resulting circles are overlaid on the original image (Figure 11D).

(A) (B)

© (D)

Figure 11: Embryo detection workflow: (A) Original Image, (B) Edge Image, (C) Hough Transform of image B and (D) Circle overlaid on the original image

Additional comments: It was observed that the instances of presence of empty embryo shell and embryo shell sized air bubbles were very less which after consulting with our client did not warrant development of code to get rid of these cases. Adult Fish detector An adult zebrafish has fully developed organs including well formed eyes which are dark black in color. This feature was used to find the location of hatched fish in images. It was observed that zebrafish eyes were at least of size 2000 pixels and minimum circularity of 0.3 at a Magnification of 1X in image feed of size 834 x 540 pixels. To detect an adult fish, an image (Figure 12A) is converted to a grayscale image and all pixels below an intensity of 50, were converted to white region and the rest were converted to black to obtain a binary image (Figure 12B). Next, all white objects which did not have a minimum circularity of 0.3 and a minimum area of 2000 pixels were discarded (Figure 12C). The image so obtained contains the eyes of adult fish in the image. These regions were then increased by dilation to highlight the upper half of an adult zebrafish(Figure 12D and 12E).

(A) (B)

(C) (D)

(E)

Figure 12: Workflow of adult fish detector: (A) Original image, (B) Thresholded image, (C) Adult Fish eye regions, (D) Dilated eye regions, (E) Overlaid results

Conclusion The detectors developed using ImageJ for the project were used as part of a larger integrated system. The detectors provided the X,Y coordinates of the fish and embryo in the slide which were then used to highlight them in the display. The highlighted objects are expected to increase the user’s involvement and learning from the system. Using the two detectors for embryo detector and adult fish detector, we were able to get good and acceptable precision and recall as can be seen in the following video:- https://www.youtube.com/watch?v=xL1TlNUJ_Gs Source Code Repository The source code for the repository can be obtained at:- https://github.com/uw-loci/exploratorium.git Detailed Analysis A detailed analysis of the performance of developed detectors on individual images can be found at:- https://github.com/uw-loci/exploratorium/blob/master/test_results.pdf

Chapter 3 Curvelet Transform and Fiber Extraction (CT FIRE) Modules

The extracellular collagen matrix (ECM) has been found to be an indicator of progression of many types of cancer. The mechanism of interaction of collagen matrix and cancer cells has been a subject of intense medical research. Development of Second Harmonic Generation (SHG) Imaging technique has enabled acquisition of high resolution images of individual fibers and subsequent analysis. For example, Conklin et al. [5] showed that patterns of arrangement of collagen fibers can predict outcome of a breast cancer patient. Nadiarnykh et al [6] discovered that patterns of collagen fibers in SHG image provided quantitative discrimination between benign and tumor tissues. Consequently, there was a need to develop an automated tool to extract individual fiber properties like fiber angle, length, straightness, width etc, from SHG images. Consequently, researchers at Laboratory of Optical and computational Instrumentation (LOCI) at UW Madison developed CT-FIRE (Curvelet Transform and Fiber Extraction Algorithm). CT-FIRE is a software tool developed in MATLAB, that allows user to automatically extract collagen fibers in an image and quantify fibers with descriptive statistics, such as fiber angle, length, straightness, and width. The program reads in image files supported by MATLAB and extracts the individual collagen fibers via a combined method we call CT-FIRE (named ctFIRE in the software implementation). The approach of CT-FIRE is described in a publication [1] ,which combines the advantage of the fast discrete curvelet transform [2] (CT) for denoising the image and enhancing the fiber edge features and the advantage of fiber extraction (FIRE) algorithm [3] for extracting individual fibers. The output may be displayed on the screen and written into .csv, .xlsx, .mat data file or .tif image files for further analysis, such as statistical test or image classification. This tool does not require prior experience of programming or image processing and can handle multiple files, enabling efficient quantification of collagen organization from biological datasets. Prior to the start of this thesis, the CT-FIRE functionality of extracting fibers from an image, existed in form of a MATLAB software. However, there was a lack of analysis tools to convert the extracted information to meaningful data. The thesis involved developing 2 main tools:-

1. Advanced Output Module 2. Region of Interest (ROI) Manager

The above tools were developed in MATLAB programming environment since the basic functionality of fiber extraction (CT-FIRE) was already developed in MATLAB.

A B C Figure 13 (A): main GUI of CT-FIRE to extract individual fiber properties including fiber length, angle, width, and straightness; (B): advanced output GUI of CT-FIRE to combine the results from multiple images with or without thresholding; and (C): region of interest (ROI) manager of CT-FIRE to do ROI annotation and analysis for a single image. Advanced Output Module This module is used in the analysis of fiber data generated by the CT-FIRE program. This module can be used to:-

1. Remove fibers from analysis 2. Visualise Fibers on image 3. Select fibers on a combination of conditions 4. Generate statistics on selected fibers 5. Visualise fibers color coded by properties

This module plots fibers with a randomly generated color along with fiber number (Figure 14). During the analysis process,it is useful to be able to visualize individual fibers plotted in the image. Individual fiber number can thus be specified for visualisation.

Figure 14: Fiber Post Processing Module and overlaid fibers

Users also have an option to remove fibers from their analysis. Users can further select fibers on a combination of conditions on 4 properties of length, width, straightness and angle. This selection can be done in either comparative terms (percentage) or in absolute terms of (pixels and degrees). After thresholding, statistics on fiber properties are generated and individual fiber properties are saved in a .xls file as shown in figures (Figure 15 A and 15 B) below .

(A) (B)

Figure 15: (A) Fiber statistics and (B) Individual fiber properties

Region of Interest (ROI) Manager Region of Interest in an image is an area with interesting features which a user is interested in analyzing. A Region of Interest allows a user to focus on a specific area and dedicate computation time and resources to the ROI instead of the entire image, which is especially useful when working with large stitched images. The ROI Manager module contains the following components:-

● Filename: displayed on the top left corner of the GUI.

● ROI table: Panel on the left contains the names of ROIs associated with the image.

● New ROI? : Dropdown menu to select the shape of ROI to be drawn. The shapes included are – rectangle, freehand, ellipse and polygon and specify. Specify option allows the user to place a rectangular ROI of specified dimensions at the specific position.

● Save ROI(s) : Saves the recently drawn ROI . ● Combine ROIs: combine multiple ROIs into one

ROI. (multiple selection of ROIs can be done by pressing “Ctrl” key and selecting the new ROI. ROIs in series can be selecting the first ROI, pressing shift and selecting the last ROI.)

● Rename ROI: change ROI name. ● Delete ROI: delete the selected ROIs. ● Measure ROI: for the selected ROIs, calculate minimum and maximum pixel values in the

ROI , Area of each ROI (units – pixels) and mean pixel value. ● Load ROI from text: allow the user to load a ROI coordinates saved in a .csv file. ● Load ROI Mask: allow the user to load a ROI saved as a mask image. ● Save ROI Text: save the selected ROI coordinates into a .csv file. ● Save ROI Mask: save the selected ROI into a .tif mask image file. ● ctFIRE ROI Analyser: Launch an advanced ROI operations window. ● Apply ctFIRE on ROI: apply ctFIRE on the selected ROI(s) to extract fibers within

them. ● Show All : display all ROIs on the image. ● Labels: display the name of ROIs along with the boundaries. Combined ROI is not labeled. ● Message Window: provide ongoing message.

Some of the functions are displayed below:-

Figure 16: ROIs plotted on image

Figure 17: Fibers in defined ROI with histogram of properties

A detailed description of the tools developed for the CT-FIRE software can be found in [4].

Conclusion and Future Direction

This work is based on multiple projects which helped me gain a great breadth of knowledge in Image Processing, Computer Vision, Data Analysis and Machine Learning. However, as with any project there is a room of improvement in each. Current fiber analysis tools include Curvelet Transform - Fiber Extraction algorithm(CT-FIRE) and CurveAlign which have different workflows and different interfaces. A future direction for could be to integrate all fiber analysis tools into one intuitive and easy to understand Graphical User Interface. The community of researchers who use the fiber analysis tools can be leveraged to improve these tools, which will reduce the development cycle. Another major push of LOCI lab is to build a smart microscope. Current imaging systems do not leverage the advances in hardware like Graphical Processing Unit (GPU) or Machine Learning during image acquisition. A considerable time is spent to acquire a stitched image of entire sample which is then analyzed by other tools. A large size image warrants a considerable time to acquire and process the image. Most of the time, a sample contains interesting phenomenon viz tumor cells in a small portion, due to which time can be saved if only those regions are acquired. Machine Learning can be used to train the imaging systems to capture regions with interesting features. Currently, researchers at LOCI are working on a project to use low resolution images which are quickly captured to find out regions with features of interest. These regions will then be imaged using a high resolution lens. It is hoped that this work will help the reader to not only tackle similar imaging problems but also serve as a guide on how to develop framework for imaging solutions

References:- 1. Bredfeldt JS, Liu Y, Pehlke CA, et al (2014) Computational segmentation of collagen fibers from second-harmonic generation images of breast cancer. J Biomed Opt 19:16007–16007. 2. Candes E, Demanet L, Donoho D, Ying L (2006) Fast discrete curvelet transforms. Multiscale Model Simul 5:861–899. 3. Stein AM, Vader DA, Jawerth LM, et al (2008) An algorithm for extracting the network geometry of three-dimensional collagen gels. J Microsc 232:463–475. doi: 10.1111/j.1365-2818.2008.02141.x 4. Bredfeldt JS, Liu Y, Mehta Gu, Keikhosravi Ad, Manual for CT-FIRE v2.0 beta http://loci.wisc.edu/files/loci/software/CT-FIRE%20V1.3%20Beta2%20Users%20Manual_0.pdf (Also attached at end of thesis) 5. Conklin M. W., et al. , “Aligned collagen is a prognostic signature for survival in human breast carcinoma,” Am. J. Pathol. 178(3), 1221–1232 (2011).http://dx.doi.org/10.1016/j.ajpath.2010.11.076[PMC free article] [PubMed] 6. Nadiarnykh O., et al. , “Alterations of the extracellular matrix in ovarian cancer studied by second harmonic generation imaging microscopy,” BMC Cancer 10(94), 1–14 (2010).http://dx.doi.org/10.1186/1471-2407-10-94[PMC free article] [PubMed] 7. Liu, Y., Keikhosravi, A., Mehta, G.S., Drifka, C.R., and Eliceiri, K.W. (2017). Methods for quantifying fibrillar collagen alignment. In Fibrosis: Methods and Protocols, L. Rittié, ed. (New York: Springer) (in press)

Appendix A(Autofocus during image acquisition) #include <stdio.h> #include <opencv2/opencv.hpp> using namespace cv; using namespace std; float main(int argc, char** argv ) { //argument check if ( argc != 2 ) { printf("usage: DisplayImage.out <Image_Path>\n"); return -1; } //image reading Mat image; image = imread( argv[1], 1 ); if ( !image.data ) { printf("No image data \n"); return -1; } //channels and rows and columns int nChannels=image.channels(); int nRows=image.rows; int nCols=image.cols; long entropy=0,fx,fy; for( int y = 1; y < image.rows-1; y++ ) { for( int x = 1; x < image.cols-1; x++ ) { for( int c = 0; c < nChannels; c++ ) { fx=(2*(long)image.at<Vec3b>(y,x)[c]- (long)image.at<Vec3b>(y,x-1)[c]- (long)image.at<Vec3b>(y,x+1)[c]); fy=(2*(long)image.at<Vec3b>(y,x)[c]- (long)image.at<Vec3b>(y-1,x)[c]- (long)image.at<Vec3b>(y+1,x)[c]); entropy=entropy+fx*fx+fy*fy; } } } float avg_entropy=(float)entropy/(nRows*nCols); cout<<argv[1]<<" entropy="<<avg_entropy<<endl; return entropy; }

Appendix B (Flat Field Correction) #include <stdio.h> #include <opencv2/opencv.hpp> using namespace cv; using namespace std; int main(int argc, char** argv ) { /* This function takes in a backgroundImage and a rawImage. Finds the mean value of the backgroundImage. Finds the field corrected image as = rawImage/backgroundImage*mean_value. The corrected image is saved as corrected_<raw_image_name> */ if(argc!=3){ printf("Correct Usage is DisplayImage.out <BackgroundImage_Path> <Raw_Image_Path>"); return -1; } Mat backgroundImage=imread(argv[1],1); Mat rawImage=imread(argv[2],1); if(!rawImage.data||!backgroundImage.data){ printf("No image data in either backgroundImage or rawImage"); return -1; } //displaying images namedWindow("Background Image", WINDOW_AUTOSIZE ); imshow("Background Image", backgroundImage); //waitKey(0); namedWindow("Raw Image", WINDOW_AUTOSIZE ); imshow("Raw Image", rawImage); waitKey(0); //error check- both images have same dimensions if(backgroundImage.channels()!=rawImage.channels()||backgroundImage.rows!=rawImage.rows||backgroundImage.cols!=rawImage.cols){ printf("raw and background images have different dimensions"); return -1; } //storing image dimensions int nChannels=rawImage.channels(); int nRows=rawImage.rows;

int nCols=rawImage.cols; //find mean value of backgroundImage long meanBackgroundValue=0; for( int c = 0; c < nChannels; c++ ){ for( int y = 0; y < nRows; y++ ){ for( int x = 0; x < nCols; x++ ){ meanBackgroundValue=meanBackgroundValue+backgroundImage.at<Vec3b>(y,x)[c]; } } } meanBackgroundValue/=nRows*nCols*nChannels; //printf("mean value=%li\n",meanBackgroundValue); for( int c = 0; c < nChannels; c++ ){ for( int y = 0; y < nRows; y++ ){ for( int x = 0; x < nCols; x++ ){ rawImage.at<Vec3b>(y,x)[c]=(int)((float)rawImage.at<Vec3b>(y,x)[c]/(float)backgroundImage.at<Vec3b>(y,x)[c]*(float)meanBackgroundValue); } } } namedWindow("Raw Image", WINDOW_AUTOSIZE ); imshow("Raw Image", rawImage); waitKey(0); //Need to save the image too char correctedImagename[200]; strcpy(correctedImagename,"corrected_"); strcat(correctedImagename,argv[2]); imwrite(correctedImagename,rawImage); return 0; }

Journal of Histochemistry & Cytochemistry 2016, Vol. 64(9) 519 –529© 2016 The Histochemical SocietyReprints and permissions: sagepub.com/journalsPermissions.navDOI: 10.1369/0022155416659249jhc.sagepub.com

Jour

nal o

f H

isto

chem

istr

y an

d C

ytoc

hem

istr

y

Article

Introduction

Basic research and clinical evaluation of cancer tissue focuses primarily on identification of malignant epithe-lial cells; however, the importance of the stroma is now appreciated as a key mediator of carcinogenesis.1,2 A major component of the tumor-associated stroma, fibril-lar collagen, has been shown to impact cancer cell behavior through biochemical and biophysical cues. Collagen normally provides the underlying structural

framework of homeostatic tissues, but alterations in specific properties such as alignment have been observed in cancer and other pathological processes.

659249 JHCXXX10.1369/0022155416659249Drifka et al.SHG and Picrosirius Red Morphometric Comparisonresearch-article2016

Received for publication May 23, 2016; accepted June 21, 2016.

Corresponding Author:Kevin W. Eliceiri, University of Wisconsin–Madison, 271 Animal Science Building, 1675 Observatory Drive, Madison, WI 53706, USA. E-mail: eliceiri@wisc.edu

Comparison of Picrosirius Red Staining With Second Harmonic Generation Imaging for the Quantification of Clinically Relevant Collagen Fiber Features in Histopathology Samples

Cole R. Drifka, Agnes G. Loeffler, Kara Mathewson, Guneet Mehta, Adib Keikhosravi, Yuming Liu, Stephanie Lemancik, William A. Ricke, Sharon M. Weber, W. John Kao, and Kevin W. EliceiriDepartment of Biomedical Engineering (CRD, AK, WJK, KWE), Laboratory for Optical and Computational Instrumentation (LOCI) (CRD, KM, GM, AK, YL, WJK, KWE), Department of Surgical Pathology (AGL), Department of Urology (SL, WAR), Department of Surgery (SMW, WJK), University of Wisconsin–Madison, Madison, Wisconsin; George M. O’Brien Research Center of Excellence, Madison, Wisconsin (WAR); University of Wisconsin Comprehensive Carbone Cancer Center, Madison, Wisconsin (AGL, WAR, SMW, WJK, KWE); and Morgridge Institute for Research, Madison, Wisconsin (CRD, KWE)

SummaryStromal collagen alignment has been shown to have clinical significance in a variety of cancers and in other diseases accompanied by fibrosis. While much of the biological and clinical importance of collagen changes has been demonstrated using second harmonic generation (SHG) imaging in experimental settings, implementation into routine clinical pathology practice is currently prohibitive. To translate the assessment of collagen organization into routine pathology workflow, a surrogate visualization method needs to be examined. The objective of the present study was to quantitatively compare collagen metrics generated from SHG microscopy and commonly available picrosirius red stain with standard polarization microscopy (PSR-POL). Each technique was quantitatively compared with established image segmentation and fiber tracking algorithms using human pancreatic cancer as a model, which is characterized by a pronounced stroma with reorganized collagen fibers. Importantly, PSR-POL produced similar quantitative trends for most collagen metrics in benign and cancerous tissues as measured by SHG. We found it notable that PSR-POL detects higher fiber counts, alignment, length, straightness, and width compared with SHG imaging but still correlates well with SHG results. PSR-POL may provide sufficient and additional information in a conventional clinical pathology laboratory for certain types of collagen quantification. (J Histochem Cytochem 64:519–529, 2016)

Keywordsfibrillar collagen, histopathology, picrosirius red, polarized microscopy, second harmonic generation

520 Drifka et al.

For example, researchers have identified a unique “tumor-associated collagen signature” (TACS) in breast cancer defined as bundles of straightened, aligned col-lagen fibers that are oriented perpendicular to the tumor boundary.3 Mechanistically, it is hypothesized that aligned collagen fibers form pathways that facilitate cancer cell migration away from the tumor and toward vasculature during the metastatic process.3,4 While there is significant research work ongoing on the role of collagen in live, dynamic in vitro and in vivo animal models of a wide array of cancers,5–10 many groups, including ours, are also focused on investigating the potential clinical utility of collagen properties in routinely fixed and processed human tissues. Toward clinical translation, it has already been shown that the detection of TACS in routine histopathological evaluation of breast cancer can serve as an optical biomarker and be pre-dictive of disease recurrence and patient survival.11 The clinical importance of collagen organization has also been investigated in pancreatic,12 esophageal,13 and ovarian cancers.14 Accordingly, there is great interest by basic and clinical researchers alike in examining and quantifying properties of stromal collagen fibers such as overall alignment as well as individual metrics includ-ing length, width, and straightness.

Many alternative methods exist to visualize collagen in the context of routine histopathological sections. For example, established histology dyes (Masson’s tri-chrome, Movat’s pentachrome, picrosirius red [PSR]),15–18 antibody detection,19 polarization-based widefield imag-ing,20 and second harmonic generation (SHG) imag-ing21,22 have all been employed. Other imaging methods often associated with collagen imaging either do not offer the appropriate resolution for the assessment of collagen organization, such as optical coherence tomography,23,24 or require non-routine sample prepara-tion methods and offer limited field of view, such as transmission electron microscopy,25 scanning electron microscopy,26 and atomic force microscopy.27 Of the imaging methods that are compatible with histopatho-logical samples, SHG is highly specific to the non-cen-trosymmetric structure of fibrillar collagen.28 SHG has become a particular focus for this reason and a variety of others, including high resolution, good signal-to-noise ratio, ability to work nondestructively on stained and unstained tissues, and power to extract additional information such as polarization properties and for-ward–backward scattering ratios. However, there is con-cern over the distinction whether SHG can be used for routine collagen visualization and quantification in more routine pathology tissues versus whether it is required. Undoubtedly, SHG has great advantages in terms of imaging depth and higher-order information, but for most of the reported histopathological studies done

with standard 5-µm-thick pathology sections, these advantages are not of primary interest.11–13,29 In addi-tion, SHG is not readily accessible in the clinical setting due to high cost, lack of clinical interfaces, and relatively slow acquisition workflows.

To this end, we wanted to determine whether a more accessible technique could provide similar quan-titative information about collagen alignment and fiber structure in histopathological samples as to what SHG can provide. Collagen fibers are composed of mole-cules that are hierarchically ordered in a parallel orien-tation resulting in inherent birefringent properties and also contain a large proportion of basic amino acid residues. Catering to both of these properties, sirius red is an elongated, acidic dye molecule that efficiently aligns and binds parallel to collagen fibers, thereby enhancing its innate birefringence. Due to its ability to specifically visualize collagen, as well as its ready availability in many research and pathology laborato-ries, PSR was chosen as a potential candidate. In this study, we systematically imaged collagen in tissue regions from 92 pancreatic cancer patients using PSR with a standard polarization widefield microscope fol-lowed by subsequent SHG imaging and then quanti-fied collagen properties from both sets of data using well-established software tools for collagen quantifica-tion.29,30 Pancreatic cancer was decided on as a ratio-nal cancer model because it remains one of the most deadly human malignancies and has been the subject of recent studies correlating collagen organization to disease progression.31 In addition, in the context of the tumor microenvironment, an extreme collagen-rich stromal response is a hallmark of pancreatic cancer and can account for up to 90% of the tumor mass.32 Furthermore, characteristic changes in collagen align-ment have been documented,12 and other stromal cues have been correlated to patient prognosis.33 Consistent with previous SHG trends,12 we found that key collagen properties, such as alignment, were com-puted to be higher in pancreatic cancer tissues com-pared with normal adjacent tissues using picrosirius red stain with standard polarization microscopy (PSR-POL). Interestingly, all of the collagen measurements that we considered were computed to be significantly higher in the PSR-POL images relative to SHG images. While the exact reasons underlying the differences remain to be determined, our data suggest that PSR-POL can extract similar, if not additional and comple-mentary, information about collagen compared with SHG. Due to a relatively low clinical adaptability bar-rier, PSR could serve as a viable alternative technique for interrogating and quantifying certain collagen prop-erties in histopathological tissues in conjunction with established fiber tracking algorithms.

SHG and Picrosirius Red Morphometric Comparison 521

Materials and Methods

Patient Tissue Staining With Picrosirius Red

To quantitatively compare collagen in multiple patients, a tissue microarray (TMA)34 available at the University of Wisconsin Comprehensive Carbone Cancer Center was utilized. This TMA was constructed (IRB No. 2011-0840-CR003) from pancreatic cancer surgical tissues acquired at the University of Wisconsin Hospital and Clinics between 1987 and 2012. All patients under-went curative-intent resection and were not treated in the neoadjuvant setting. Hematoxylin and eosin (H&E) sections were prepared from formalin-fixed paraffin embedded (FFPE) blocks originally used for diagno-sis and reviewed by board-certified pathologists to identify different representative tumor regions. Blocks were then punched using a 1-mm-diameter core and inserted into a recipient FFPE block using a Beecher Instruments (Sun Prairie, WI) MTA-1 manual tissue arrayer. 5-μm-thick sections were cut from the TMA blocks on a microtome and adhered to standard glass slides. A PSR solution was formulated by dissolving 0.1-g Fast Green FCF (F7258; Sigma-Aldrich, St. Louis, MO) and 0.1-g Sirius Red F3B (S03695; Pfaltz & Bauer, Waterbury, CT) in 100 ml of saturated picric acid. After deparaffinization and rehydration, the TMA sections were covered with one to two drops of the PSR solution for 60 min at 40C. The slides were then rinsed with acidified water, dehydrated through three changes of 100% ethanol and three changes of xylene, and coverslipped with toluene-based Richard-Allan mounting medium (Thermo Fisher Scientific, Waltham, MA). Representative human breast and prostate cancer tissues were also stained using the same protocol.

Picrosirius Red Microscopy

The PSR-stained tissues were digitalized using an Aperio slide scanner (Leica Biosystems, Vista, CA) to enable easy pathology review for region of interest (ROI) selection. For the pancreatic cancer TMA, two to three ROIs containing cancer cells and associated stroma were annotated on 92 patient tumors. All PSR imaging was performed on a traditional transmission pathology microscope (BX53; Olympus Corp.) using a 20× objec-tive (UPlanFL N NA = 0.50; Olympus Corp.) without polarizers (PSR) and with linear polarizers (PSR-POL). Halogen lamp intensity was kept constant for both image types, and an exposure time that optimized the signal-to-noise ratio was chosen and kept constant within each image type. All images were digitally cap-tured using cellSens acquisition platform (Olympus Corp.). To control for intrinsic tissue birefringence, we

also imaged unstained pancreatic tissue without and with polarized light (Supplementary Fig. S1).

Second Harmonic Generation Microscopy

The same ROIs were imaged using a custom-built SHG microscope developed by the Laboratory for Optical and Computational Instrumentation known as CAMM (Compact Automated Multiphoton Microscope). The CAMM was designed to screen histology samples for cancer-associated changes in collagen fiber prop-erties.29 Tissue landmarks were used to align the field of view and assure the exact same ROIs were imaged as on the PSR-stained sections. SHG signal was gen-erated using 780-nm excitation light from a Ti:sapphire femtosecond laser and detected in the forward direc-tion after spectrally filtering (390/18 BP; Semrock, Inc., Rochester, NY) using a 20× objective (S Fluor NA = 0.75; Nikon Corp.). All images were acquired at 1024 × 1024 resolution using WiscScan software (http://loci.wisc.edu/software/wiscscan) with consistent parame-ters. Fifteen of the pancreatic cancer ROIs were also imaged using a backward-detection SHG microscope. For this, the SHG signal was generated using 890-nm excitation light and collected with the same objective after filtering (445/20 BP; Semrock, Inc.).

Fiber Segmentation and Measurement

PSR-POL images were converted to 8-bit, cropped, and downsized in the ImageJ-based analysis pack-age Fiji35 to align with the field of view and pixel res-olution of the SHG images. CT-FIRE fiber analysis software30 (http://loci.wisc.edu/software/ctfire) was then applied to the SHG and PSR-POL images using default settings after thresholding background (20–255 for SHG, 5–255 for PSR-POL). Quantitative information computed by CT-FIRE for each image type included the individual fiber metrics of length, width, and straightness. Collective fiber properties of alignment and total count were also computed for the entire image.

Statistics

For every CT-FIRE metric, a patient value was calcu-lated by averaging the mean values for the two to three ROIs. All data were verified to follow Gaussian sam-pling using the Shapiro–Wilk normality test. Values obtained using SHG and PSR-POL were compared using a repeated-measures paired t-test and Pearson correlation analysis in GraphPad Prism (GraphPad Software, Inc., La Jolla, CA). A value of p<0.05 (two-tailed) was considered statistically significant.

522 Drifka et al.

Fig

ure

1. P

anel

of c

omm

only

use

d hi

stol

ogic

al d

yes

and

mic

rosc

opy

met

hods

to

visu

aliz

e co

llage

n co

mpl

emen

tary

to

stan

dard

hem

atox

ylin

and

eos

in s

tain

ing

in t

hin

sect

ions

. A

lso

show

n is

sec

ond

harm

onic

gen

erat

ion,

whi

ch r

emai

ns t

he e

xper

imen

tal “

gold

sta

ndar

d.”

All

stai

ning

was

don

e on

ser

ial s

ectio

ns o

f pan

crea

tic c

ance

r tis

sue.

Sca

le b

ars

=

100

µm.

SHG and Picrosirius Red Morphometric Comparison 523

Fig

ure

2. R

epre

sent

ativ

e hu

man

pan

crea

tic t

issu

e vi

sual

ized

with

PSR

-PO

L an

d SH

G. U

nder

bri

ghtf

ield

tra

nsm

issi

on m

icro

scop

y, P

SR s

tain

s co

llage

ns w

ith a

dee

p re

d co

lor.

W

hen

pair

ed w

ith p

olar

izat

ion

mic

rosc

opy,

PSR

ena

bles

vis

ualiz

atio

n of

hig

hly

bire

frin

gent

thi

ck (

oran

ge-r

ed)

and

thin

(gr

een-

yello

w)

colla

gen

fiber

s. C

olla

gen

orga

niza

tion

has

prev

ious

ly b

een

show

n to

be

char

acte

rist

ical

ly a

ligne

d in

the

panc

reat

ic c

ance

r st

rom

a us

ing

SHG

. Abb

revi

atio

ns: P

SR-P

OL,

pic

rosi

rius

red

sta

in w

ith s

tand

ard

pola

riza

tion

mic

ros-

copy

; SH

G, s

econ

d ha

rmon

ic g

ener

atio

n; P

SR, p

icro

siri

us r

ed. S

cale

bar

s =

100

µm

.

524 Drifka et al.

Results and Discussion

Histological examination of biopsied or resected tissue by surgical pathology remains the current “gold stan-dard” for cancer diagnosis, staging, and patient prog-nostication. In practice, several large tissue sections are routinely examined, but only the epithelial ele-ments are taken into consideration in determining grade and stage of the tumor, and consequently, only these inform decisions regarding clinical manage-ment. Collagen has long been qualitatively appreci-ated as a mainstay tissue component of tumors, but its features have yet to be incorporated into pathologists’ overall assessment. With the burgeoning evidence for collagen structure and organization as important players in progression of multiple cancer types, there is growing interest by pathologists to integrate quanti-tative information about collagen and other features of the tumor microenvironment with traditional qualitative observations to provide a more nuanced characteriza-tion of the tissue under examination.

To date, basic research of collagen has been greatly accelerated by SHG imaging, a laser scanning

microscopy technique that can provide high-resolution, quantifiable images of discrete collagen fibers without the need for exogenous staining. However, the technology currently is poorly suited for pathology workflows due to an array of issues including expense, instrumentation complexity and footprint, technically demanding opera-tion, and clinical acceptance. Furthermore, collagen assessment needs further clinical validation. To adopt assessment of collagen structure and organization to pathological practice, a surrogate technique that provides certain types of collagen quantification information similar to that of SHG imaging and can be used routinely in the pathology laboratory needs to be identified. Although H&E is the most common histological stain used in clini-cal pathological practice, several complementary dyes are routinely used in surgical pathology labs that are capable of visualizing stromal collagen (Fig. 1). Masson’s trichrome and Movat’s pentachrome are used to visualize connective tissue on the basis of color, but their specificity to collagen is limited. PSR also differentially stains con-nective tissue with a rich red color, and it can be com-bined with conventional widefield polarization microscopy to greatly enhance the birefringence of collagen fibers

Figure 3. Quantification of different collagen fiber metrics (alignment, length, straightness, and width) in normal and pancreatic cancer-associated stroma using picrosirius red stain with standard polarization microscopy imaging. Images were acquired from matched normal and cancer tissues from 75 patients.

SHG and Picrosirius Red Morphometric Comparison 525

(Supplementary Fig. S1) and specifically visualize them with a green–yellow–orange–red spectrum.36 PSR has already been exploited to understand collagen changes in cancer and other pathologies associated with fibro-sis.16,18,37–39 Consistent with our previous observations,12 stromal collagen appears more elongated and aligned in the immediate stroma surrounding malignant ducts com-pared with normal pancreas ducts when examined using PSR-POL (Fig. 2). Highly remodeled collagen was also observed in breast (Supplementary Fig. S2)11 and prostate (Supplementary Fig. S3)38 cancers using PSR-POL.

To validate the utility of PSR-POL to detect collagen changes, we acquired images around histologically confirmed normal and malignant pancreatic ducts. Using SHG, we previously established that collagen is characteristically reorganized and shows increased alignment, length, straightness, and width around malignant pancreatic ducts compared with benign ducts.12 Here, we determined that PSR-POL imaging yields similar data: collagen fibers are more aligned, elongated, and thick in the stroma adjacent to the malignant epithelium (Fig. 3, p<0.0001). Fiber straight-ness was not significantly different between normal and cancer stroma when examined with PSR-POL (p=0.4391).

We next quantitatively compared images acquired in the same ROIs of 92 pancreatic cancer patient tumors using PSR-POL and SHG. For each technique, image acquisition parameters were set, so that the best signal-to-noise ratio was obtained and images were therefore representative of optimal visualization capabilities. We decided to employ forward-detection SHG in this com-parison study due to our experience in achieving brighter signals in thin samples compared with backward-detec-tion SHG (Supplementary Fig. S4).40 We also quantita-tively confirmed that forward- and backward-detection SHG configurations yield statistically comparable fiber measurements (Supplementary Fig. S5). Although many of the same collagen features are seen by both PSR-POL and SHG, it was apparent that the images did not perfectly mirror each other (Figs. 2, S2, and S3). We quantified the images using an open-source, validated software program, CT-FIRE, developed by our group, which is capable of reducing noise, extracting, and mea-suring important fiber metrics, including count, align-ment, length, straightness, and width.30 The quantitative scoring ability of CT-FIRE has several advantages over standard qualitative examination including increased objectivity, reproducibility, and sensitivity. We observed that CT-FIRE was compatible with grayscale images generated by both SHG and PSR-POL modalities (Supplementary Fig. S4). As depicted in Fig. 4, we deter-mined that PSR-POL produced higher fiber count, length, width, and straightness values than SHG

Figure 4. Left column: Quantification of different collagen fiber metrics (count, alignment, length, straightness, and width) in images acquired from pancreatic cancer tissues using SHG and PSR-POL. Right column: Pearson correlation analysis between metrics generated by SHG and PSR-POL imaging. Data represent 92 patient tumors. Abbreviations: SHG, second harmonic gen-eration; PSR-POL, picrosirius red stain with standard polarization microscopy.

526 Drifka et al.

(p<0.0001). Alignment was also higher in PSR-POL images (p=0.0154). Similar quantitative trends were observed in prostate tissues (Supplementary Fig. S6), suggesting that the results were not cancer-type spe-cific. Using Pearson correlation analysis, we determined that length (r = 0.159, p=0.129) and width (r = 0.174, p=0.097) were not significantly correlated between the two techniques. Straightness (r = 0.430, p<0.0001), alignment (r = 0.565, p<0.0001), and count (r = 0.656, p<0.0001) did show significant positive correlation between PSR-POL and SHG.

Based on our data, we believe that PSR-POL may be a practical option to investigate stromal collagen in thin tissue sections depending on the application and particular metric under examination. For applications that require absolute measurements predetermined by SHG, PSR-POL needs to be considered carefully because the technique appears to generate statisti-cally higher values than SHG. The reasons for these differences are currently unknown. It has been suggested that PSR-POL can be sensitive to fiber thickness,16 maturity,18 and collagen type.37 Further investigations will be of great interest to determine which collagen types and characteristics are being visualized or missed by each imaging technique.41 However, for applications comparing the relative mag-nitudes of fiber properties, PSR-POL is positively cor-related to SHG for count, alignment, and straightness and therefore might be a viable alternative for assess-ing these in routine pathological practice. Of particular interest is fiber alignment, which out of all collagen metrics assessed here has been demonstrated to cor-relate most strongly with disease progression and yields slightly higher measurements in PSR-POL

images compared with SHG. As the emerging clinical value of other fiber metrics (ie, length, width) become clearer, imaging methods and technologies will likely need to be reconsidered to determine which would be most appropriate for facile visualization and analysis.

There are advantages and disadvantages to SHG and PSR-POL as outlined in Table 1. SHG is optimal for directly visualizing fibrillar collagens42 with submi-cron resolution; however, its use mandates costly equipment and technical expertise. On the other hand, PSR-POL is already utilized in many histology labs and therefore has a low cost and barrier to implemen-tation. Although PSR dye alone does not specifically stain collagen,43 it can be used to directly visualize anatomic structures within tissue sections, which is important for spatial orientation and analysis of stro-mal collagen relative to a defined tumor boundary, as demonstrated in breast cancer.3,29 When PSR is paired with polarized light, collagen can then be visualized in the same field of view due to specific enhancement in birefringence by certain types of fibrillar collagen.41 PSR-POL can also be combined with fluorescence microscopy44 or more advanced polarimetry setups45 to gain additional information about the tissue. Although not encountered in this study, PSR-POL may also have limitations including potential reproducibility issues due to variations in staining protocols46 and will need to be assessed in other tissues. Until further vali-dation studies are performed, SHG remains the gold standard for assessing collagen alignment in the research setting, particularly for studies requiring deep three-dimensional intravital imaging,4 forward–back-ward scattering ratios,40 and circular versus linear polarization dependencies.47

Table 1. Advantages and Disadvantages of Each Imaging Modality for Collagen Visualization.

SHG PSR-POL

Contrast mechanism

Intrinsic signal from light interaction with non-centrosymmetric fibrillar collagen structure

Collagen affinity for Sirius Red F3B

Specificity Direct visualization of fibrillar collagens (ie, types I, III). Also sensitive to other harmonophores (actomyosin, microtubules)

PSR dye alone is not specific for collagen, but when paired with polarized light, fibrillar collagens undergo specific enhancement in birefringence enabling visualization

Equipment Extensive equipment Standard transmission pathology scope with two accurately crossed polarization filters (polarizer and analyzer)

Cost Expensive InexpensiveResolution Easier to visualize individual collagen fibers. Forward–

backward scattering ratioEasier to visualize bundles of collagen. Detected

straightness and width significantly higherUtility No staining required and can be done on unlabeled

sections making retrospective studies possible. Submicron resolution allows visualization of individual fibers. Forward–backward scattering ratio. Can optically section thick (500–1000 µm) tissues

Requires staining of sections which might not be possible for retrospective studies. Resolution limited by visible light illumination. Can visualize counterstained cell structures directly without polarizer in place enabling examination of the stroma relative to a boundary

Abbreviations: SHG, second harmonic generation; PSR-POL, picrosirius red stain with standard polarization microscopy; PSR, picrosirius red.

SHG and Picrosirius Red Morphometric Comparison 527

Given accumulating literature that points to a cor-relation between collagen properties in peritumoral stroma to stage and patient outcome, there is desire by clinical and experimental pathologists to character-ize tumor tissue in a manner that includes assessment of collagen organization in the spatial context of malig-nant cells. PSR-POL is an attractive method for visual-izing collagen fibers in histopathological sections, along with routine H&E, due to its convenient availabil-ity in most clinical laboratories and ability to specifi-cally isolate collagen contrast with linear polarizers on a brightfield microscope. PSR-POL generates images that are quantifiable by existing software packages and comparable with SHG, which has so far served as the gold standard for quantitative assessment of col-lagen. PSR-POL could therefore be a practical, easy-to-implement method for improving the evaluation of cancerous tissues in the clinical pathology setting.

Acknowledgments

We thank Mina Alfi for useful discussions, Dr Ruth Sullivan for pathology advice, Dr Lee Wilke and Dr Patricia Keely for access to human breast cancer tissues, and Peter Crump for guidance on statistical analyses.

Author Contributions

CRD, AGL, WAR, SMW, WJK, and KWE conceived the study and drafted the manuscript. AGL provided pathology guidance. CRD, KM, GM, and SL performed all of the imag-ing. AK and YL assisted in image segmentation and quanti-fication. CRD carried out all data analysis. All authors have read and approved the final manuscript.

Competing Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The tissue microarray resource used in this study was made possible by a generous donation to the University of Wisconsin Comprehensive Carbone Cancer Center by Teresa’s Foundation for Pancreatic Cancer. We thank National Institutes of Health for their financial support P30 CA014520 (University of Wisconsin Comprehensive Carbone Cancer Center facilities and services) and U54 DK104310 (to WAR and KWE).

Literature Cited

1. Cunha GR, Ricke W, Thomson A, Marker PC, Risbridger G, Hayward SW, Wang YZ, Donjacour AA, Kurita T. Hormonal, cellular, and molecular regulation of normal and

neoplastic prostatic development. J Steroid Biochem Mol Biol. 2004;92:221–36.

2. Ricke EA, Williams K, Lee YF, Couto S, Wang Y, Hayward SW, Cunha GR, Ricke WA. Androgen hormone action in prostatic carcinogenesis: stromal androgen receptors mediate prostate cancer progression, malig-nant transformation and metastasis. Carcinogenesis. 2012;33:1391–8.

3. Provenzano PP, Eliceiri KW, Campbell JM, Inman DR, White JG, Keely PJ. Collagen reorganization at the tumor-stromal interface facilitates local invasion. BMC Med. 2006;4:38.

4. Condeelis J, Segall JE. Intravital imaging of cell move-ment in tumours. Nat Rev Cancer. 2003;3:921–30.

5. Brown E, McKee T, diTomaso E, Pluen A, Seed B, Boucher Y, Jain RK. Dynamic imaging of collagen and its modulation in tumors in vivo using second-harmonic generation. Nat Med. 2003;9:796–800.

6. Perentes JY, McKee TD, Ley CD, Mathiew H, Dawson M, Padera TP, Munn LL, Jain RK, Boucher Y. In vivo imaging of extracellular matrix remodeling by tumor-associated fibroblasts. Nat Methods. 2009;6:143–5.

7. Sung KE, Yang N, Pehlke C, Keely PJ, Eliceiri KW, Friedl A, Beebe DJ. Transition to invasion in breast cancer: a microfluidic in vitro model enables exami-nation of spatial and temporal effects. Integr Biol. 2011;3:439–50.

8. Carey SP, Kraning-Rush CM, Williams RM, Reinhart-King CA. Biophysical control of invasive tumor cell behavior by extracellular matrix microarchitecture. Biomaterials. 2012;33:4157–65.

9. Kakkad SM, Solaiyappan M, Argani P, Sukumar S, Jacobs LK, Leibfritz D, Bhujwalla ZM, Glunde K. Collagen I fiber density increases in lymph node positive breast cancers: pilot study. J Biomed Opt. 2012;17:116017.

10. Drifka CR, Eliceiri KW, Weber SM, Kao WJ. A bioen-gineered heterotypic stroma-cancer microenvironment model to study pancreatic ductal adenocarcinoma. Lab Chip. 2013;13:3965–75.

11. Conklin MW, Eickhoff JC, Riching KM, Pehlke CA, Eliceiri KW, Provenzano PP, Friedl A, Keely PJ. Aligned collagen is a prognostic signature for survival in human breast carcinoma. Am J Pathol. 2011;178:1221–32.

12. Drifka CR, Tod J, Loeffler AG, Liu Y, Thomas GJ, Eliceiri KW, Kao WJ. Periductal stromal collagen topol-ogy of pancreatic ductal adenocarcinoma differs from that of normal and chronic pancreatitis. Mod Pathol. 2015;28:1470–80.

13. Hanley CJ, Noble F, Ward M, Bullock M, Drifka C, Mellone M, Manousopoulou A, Johnston HE, Hayden A, Thirdborough S, Liu Y, Smith DM, Mellows T, Kao WJ, Garbis SD, Mirnezami A, Underwood TJ, Eliceiri KW, Thomas GJ. A subset of myofibroblastic cancer-asso-ciated fibroblasts regulate collagen fiber elongation, which is prognostic in multiple cancers. Oncotarget. 2015;7:6159–74.

14. Nadiarnykh O, LaComb RB, Brewer MA, Campagnola PJ. Alterations of the extracellular matrix in ovarian

528 Drifka et al.

cancer studied by second harmonic generation imaging microscopy. BMC Cancer. 2010;10:94.

15. Ayala G, Tuxhorn JA, Wheeler TM, Frolov A, Scardino PT, Ohori M, Wheeler M, Spitler J, Rowley DR. Reactive stroma as a predictor of biochemical-free recurrence in prostate cancer. Clin Cancer Res. 2003;9:4792–801.

16. Allon I, Vered M, Buchner A, Dayan D. Stromal differ-ences in salivary gland tumors of a common histopatho-genesis but with different biological behavior: a study with picrosirius red and polarizing microscopy. Acta Histochem. 2006;108:259–64.

17. Provenzano PP, Cuevas C, Chang AE, Goel VK, Von Hoff DD, Hingorani SR. Enzymatic targeting of the stroma ablates physical barriers to treatment of pancreatic ductal adenocarcinoma. Cancer Cell. 2012;21:418–29.

18. Arun Gopinathan P, Kokila G, Jyothi M, Ananjan C, Pradeep L, Humaira Nazir S. Study of collagen bire-fringence in different grades of oral squamous cell carcinoma using picrosirius red and polarized light microscopy. Scientifica (Cairo). 2015;2015:802980.

19. Brisson BK, Mauldin EA, Lei W, Vogel LK, Power AM, Lo A, Dopkin D, Khanna C, Wells RG, Puré E, Volk SW. Type III collagen directs stromal organization and limits metastasis in a murine model of breast cancer. Am J Pathol. 2015;185:1471–86.

20. Oldenbourg R. Polarized light field microscopy: an ana-lytical method using a microlens array to simultaneously capture both conoscopic and orthoscopic views of bire-fringent objects. J Microsc. 2008;231:419–32.

21. Hompland T, Erikson A, Lindgren M, Lindmo T, de Lange Davies C. Second-harmonic generation in collagen as a potential cancer diagnostic parameter. J Biomed Opt. 2008;13:054050.

22. Campagnola PJ. Second harmonic generation imaging microscopy: applications to diseases diagnostics. Anal Chem. 2011;83:3224–31.

23. Testoni PA, Mangiavillano B. Optical coherence tomog-raphy in detection of dysplasia and cancer of the gastro-intestinal tract and bilio-pancreatic ductal system. World J Gastroenterol. 2008;14:6444–52.

24. Yi J, Radosevich AJ, Stypula-Cyrus Y, Mutyal NN, Azarin SM, Horcher E, Goldberg MJ, Bianchi LK, Bajaj S, Roy HK, Backman V. Spatially resolved optical and ultrastructural properties of colorectal and pancreatic field carcinogenesis observed by inverse spectro-scopic optical coherence tomography. J Biomed Opt. 2014;19:36013.

25. Quan BD, Sone ED. Cryo-TEM analysis of collagen fibrillar structure. Methods Enzymol. 2013;532:189–205.

26. Kobayashi M, Furuya Y, Okabayashi T, Araki K. Scanning electron microscopic study of the three-dimensional structure of the collagen sheath surround-ing cancer cells after single high-dose irradiation. Med Mol Morphol. 2006;39:106–12.

27. Plodinec M, Loparic M, Monnier CA, Obermann EC, Zanetti-Dallenbach R, Oertle P, Hyotyla JT, Aebi U, Bentires-Alj M, Lim RYH, Schoenenberger C-A.

The nanomechanical signature of breast cancer. Nat Nanotechnol. 2012;7:757–65.

28. Campagnola PJ, Millard AC, Terasaki M, Hoppe PE, Malone CJ, Mohler WA. Three-dimensional high-reso-lution second-harmonic generation imaging of endog-enous structural proteins in biological tissues. Biophys J. 2002;82:493–508.

29. Bredfeldt J, Liu Y, Conklin M, Keely P, Mackie T, Eliceiri K. Automated quantification of aligned collagen for human breast carcinoma prognosis. J Pathol Inform. 2014;5:28.

30. Bredfeldt JS, Liu Y, Pehlke CA, Conklin MW, Szulczewski JM, Inman DR, Keely PJ, Nowak RD, Mackie TR, Eliceiri KW. Computational segmentation of collagen fibers from second-harmonic generation images of breast cancer. J Biomed Opt. 2014;19:16007.

31. Hidalgo M. Pancreatic cancer. N Engl J Med. 2010;362:1605–17.

32. Erkan M, Hausmann S, Michalski CW, Fingerle AA, Dobritz M, Kleeff J, Friess H. The role of stroma in pan-creatic cancer: diagnostic and therapeutic implications. Nat Rev Gastroenterol Hepatol. 2012;9:454–67.

33. Erkan M, Michalski CW, Rieder S, Reiser-Erkan C, Abiatari I, Kolb A, Giese NA, Esposito I, Friess H, Kleeff J. The activated stroma index is a novel and indepen-dent prognostic marker in pancreatic ductal adenocarci-noma. Clin Gastroenterol Hepatol. 2008;6:1155–61.

34. Kononen J, Bubendorf L, Kallioniemi A, Bärlund M, Schraml P, Leighton S, Torhorst J, Mihatsch MJ, Sauter G, Kallioniemi OP. Tissue microarrays for high-through-put molecular profiling of tumor specimens. Nat Med. 1998;4:844–7.

35. Schindelin J, Arganda-Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T, Preibisch S, Rueden C, Saalfeld S, Schmid B, Tinevez J-Y, White DJ, Hartenstein V, Eliceiri K, Tomancak P, Cardona A. Fiji: an open-source platform for biological-image analysis. Nat Methods. 2012;9:676–82.

36. Street JM, Souza ACP, Alvarez-Prats A, Horino T, Hu X, Yuen PST, Star RA. Automated quantification of renal fibrosis with sirius red and polarization contrast micros-copy. Physiol Rep. 2014;2:e12088.

37. Zhang H, Sun L, Wang W, Ma X. Quantitative analy-sis of fibrosis formation on the microcapsule surface with the use of picro-sirius red staining, polarized light microscopy, and digital image analysis. J Biomed Mater Res A. 2006;76:120–5.

38. Bauman TM, Nicholson TM, Abler LL, Eliceiri KW, Huang W, Vezina CM, Ricke WA. Characterization of fibrillar collagens and extracellular matrix of glandu-lar benign prostatic hyperplasia nodules. PLoS ONE. 2014;9:e109102.

39. Manjunatha BS, Agrawal A, Shah V. Histopathological evaluation of collagen fibers using picrosirius red stain and polarizing microscopy in oral squamous cell carci-noma. J Cancer Res Ther. 2015;11:272–6.

40. Abraham T, Kayra D, McManus B, Scott A. Quantitative assessment of forward and backward second harmonic three dimensional images of collagen type I matrix

SHG and Picrosirius Red Morphometric Comparison 529

remodeling in a stimulated cellular environment. J Struct Biol. 2012;180:17–25.

41. Junqueira LC, Cossermelli W, Brentani R. Differential staining of collagens type I, II and III by Sirius Red and polarization microscopy. Arch Histol Jpn. 1978;41:267–74.

42. Strupler M, Pena AM, Hernest M, Tharaux P-L, Martin J-L, Beaurepaire E, Schanne-Klein M-C. Second har-monic imaging and scoring of collagen in fibrotic tissues. Opt Express. 2007;15:4054–65.

43. Junqueira LCU, Bignolas G, Brentani RR. Picrosirius staining plus polarization microscopy, a specific method for collagen detection in tissue sections. Histochem J. 1979;11:447–55.

44. Vogel B, Siebert H, Hofmann U, Frantz S. Determination of collagen content within picrosirius red stained par-affin-embedded tissue sections using fluorescence microscopy. MethodsX. 2015;2:124–34.

45. Bancelin S, Nazac A, Ibrahim BH, Dokládal P, Decencière E, Teig B, Haddad H, Fernandez H, Schanne-Klein M-C, De Martino A. Determination of collagen fiber orienta-tion in histological slides using Mueller microscopy and validation by second harmonic generation imaging. Opt Express. 2014;22:22561–74.

46. Nazac A, Bancelin S, Teig B, Ibrahim BH, Fernandez H, Schanne-Klein M-C, De Martino A. Optimization of Picrosirius red staining protocol to determine collagen fiber orientations in vaginal and uterine cervical tissues by Mueller polarized microscopy. Microsc Res Tech. 2015;78:723–30.

47. Doras C, Taupier G, Barsella A, Mager L, Boeglin A, Bulou H, Bousquet P, Dorkenoo KD. Polarization state studies in second harmonic generation sig-nals to trace atherosclerosis lesions. Opt Express. 2011;19:15062–8.