1. technology and computing
2. internet technology
3. web clip art

568996
research-article2015
JHCXXX10.1369/0022155415568996Linden et al.Method for Consistent Images
Article
Journal of Histochemistry & Cytochemistry 2015, Vol. 63(4) 233­–243
© The Author(s) 2015
Reprints and permissions:
sagepub.com/journalsPermissions.nav
DOI: 10.1369/0022155415568996
jhc.sagepub.com
An Innovative Method for Obtaining Consistent Images
and Quantification of Histochemically Stained Specimens
Michael A. Linden, Gerald J. Sedgewick, and Marna Ericson
Department of Lab Medicine and Pathology, University of Minnesota, Minneapolis, Minnesota (MAL); Imaging and Analysis, Saint Paul, Minnesota (GJS);
and Department of Dermatology, University of Minnesota, Minneapolis, Minnesota (ME)
Summary
Obtaining digital images of color brightfield microscopy is an important aspect of biomedical research and the clinical
practice of diagnostic pathology. Although the field of digital pathology has had tremendous advances in whole-slide
imaging systems, little effort has been directed toward standardizing color brightfield digital imaging to maintain imageto-image consistency and tonal linearity. Using a single camera and microscope to obtain digital images of three stains,
we show that microscope and camera systems inherently produce image-to-image variation. Moreover, we demonstrate
that post-processing with a widely used raster graphics editor software program does not completely correct for sessionto-session inconsistency. We introduce a reliable method for creating consistent images with a hardware/software solution
(ChromaCal™; Datacolor Inc., NJ) along with its features for creating color standardization, preserving linear tonal levels,
providing automated white balancing and setting automated brightness to consistent levels. The resulting image consistency
using this method will also streamline mean density and morphometry measurements, as images are easily segmented and
single thresholds can be used. We suggest that this is a superior method for color brightfield imaging, which can be used
for quantification and can be readily incorporated into workflows. (J Histochem Cytochem 63:233–243, 2015)
Keywords
color brightfield, camera, ChromaCal, density, histology, image consistency, linearity, microscope, morphometry,
quantification
Introduction
When working with color brightfield (CB) images, a great
deal of effort has been put into the improvement of camera
and whole-slide imaging (WSI) systems, but little effort has
been directed toward the development of a workflow that
increases the consistency between images, maintains tonal
linearity, and improves image quality, and does so using
good laboratory practices (GLP). We address these issues
and introduce a novel method (ChromaCal™; Datacolor
Inc., NJ), which includes standardization to a matrix of colors on a calibration slide, automated white balancing, automated brightness, tracking and reporting of post-processing,
and algorithms that preserve linear tonal levels as well as
camera linearity assessment.
Many of the problems inherent to image inconsistency
when performing CB digital photomicroscopy stem from
operator variability in using camera systems and microscopes, as well as inaccuracies in camera systems.
Optimally, a high-quality image should lack the following:
muted colors (Margulis 2006; Min 2009; Luo et al. 1991)
background discoloration (Lee 2005; Viggiano 2004),
over- and under-exposed (saturated) images (Cromey 2010;
Received for publication October 7, 2014; accepted December 10, 2014.
Corresponding Author:
Michael A. Linden, Department of Lab Medicine and Pathology,
University of Minnesota, Minneapolis, MMC 609 Mayo, 420 Delaware St
SE, Minneapolis, MN 55455, USA.
Email: Linde013@umn.edu
Downloaded from jhc.sagepub.com by guest on March 26, 2015
234
Linden et al.
Martin and Blatt 2013), and non-linearity (Chieco et al.
1994; Garcia et al. 2013; Price and Gray 2011; Rossner and
Yamada 2004). Moreover, the images should have consistent brightness levels (Inoue 1986; Sedgewick 2008).
Most image defects can be better controlled when the
user is familiar with using each camera system and microscope. The recommended steps include setting the attenuation of light to the manufacturer’s recommended setting
(when the light source can be attenuated), focusing on the
desired area of the specimen, setting Köhler illumination (a
method for collimation of light through the sample), setting
the exposure, and finding the optimal non-tissue area to use
for white balancing the camera in the camera acquisition
software (Matsumoto 2010). Even when performing all
steps properly, this paper will demonstrate that inconsistencies in images are unavoidable, and a post-processing correction is vital for high-quality images.
In this study, we discuss the major image inconsistencies
found in published images that are often less obvious than
poor focus or poor contrast. We do not discuss microscope
illumination problems, including uneven illumination
across the image (Leong et al. 2003) or chromatic aberration (false color) at edges of cells and tissues (Foster 1997).
However, we tested a microscope/camera operator with a
single camera and microscope for images acquired over
more than one microscope session using manual and automated exposure to determine levels of image inconsistency.
Starting with a suboptimal image, we evaluated post-processing procedures to improve image quality by comparing
a popularly used program (Photoshop) with the ChromaCal
system. Using the ChromaCal system to improve image
quality to consistent levels, we assessed its ability to provide predicted measurements for both optical density and
morphometric measurements, using a single threshold value
and tested against threshold values set manually by the user.
Materials & Methods
All of the following studies were reviewed by our
Institutional Review Board and deemed exempt. Unless
noted otherwise, all experiments used the following samples, equipment, software, and microscope operator.
Histological Preparation of Sample Slides
Hematoxylin and Eosin (H&E) and Immunoperoxidase (IP)
Stains. Bone marrow core biopsies were fixed in acetic acidzinc-formalin (AZF), lightly decalcified with acid, then processed for histology (dehydrate and imbed in paraffin).
Cores were mounted onto blocks and sectioned at 3 µm and
either stained with H&E or processed for automated IHC
staining (Ventana Benchmark Immunostainer; Ventana
Medical Systems, Inc., Tuscon, AZ). The anti-kappa immunostain was purchased prediluted (mouse monoclonal, clone
KDB-1; Biocare, Concord, CA). Antigen retrieval was performed using mild Cell Conditioning 1 solution (CC1, Ventana Medical Systems, Inc.). The primary antibody was
incubated for 4 min at 37°C. The anti-CD61 immunostain
was purchased prediluted (mouse monoclonal, clone 2F2;
Cell Marque, Rocklin, CA). Antigen retrieval was performed using standard CC1 solution. The primary antibody
was incubated for 32 min at 37°C. Detection of both primary
antibodies was performed using the Ultraview Detection Kit
(Ventana Medical Systems).
Wright–Giemsa (WG). Bone marrow aspirates were collected, smeared on slides and air dried. Staining was performed as previously described (Dunning and Safo 2012).
Briefly, air-dried smear slides were alcohol fixed and then
stained in a buffered WG stain. The polychromatic stain
contains methylene blue, azures (oxidative products of
methylene blue), and eosin, which differentially stain the
nuclear and cytoplasmic components.
Imaging Equipment and Software
Images were captured on an Olympus BX51 microscope
(Olympus; Tokyo, Japan) with a halogen light source and
light-balancing daylight (LBD) filter in the light path, with
attenuation set at the factory recommended setting for camera use, outfitted with plan-apochromatic, infinity corrected
lenses: 10× (NA 0.25), 20× oil (NA 0.8), 40× oil (NA 1.0)
and 100× oil (NA 1.3). Unless otherwise described, the
camera for imaging was the Dage XL MCT (Dage-MTI;
Michigan City, IN) with Exponent 7 acquisition software
(Stable Microsystems; Surrey, UK).
ChromaCal software, version 1.4 (Datacolor Inc.;
Lawrenceville, NJ) and Photoshop software, version 10
(Adobe Systems Inc.; San Jose, CA), were used to postprocess images. Photoshop was used to segment and measure images; Matlab version 2012a (32-bit; Matlab Inc.,
Natick, MA), Fiji version 1.47t (ImageJ; Wayne Rasband,
Bethesda, MD) and a MatLab-created program (RGB_DE_
Calculation D50; Hong Wei, Datacolor, Inc., Lawrenceville,
NJ) were used for image measurements.
Study 1: Image Variability due to Imaging System and Operator
Study Outline: Images from a single microscopist over multiple sessions comparing automated and manual camera settings. For this study, a microscope/camera operator set
up the microscope with fixed Köhler illumination, set the
attenuation as described above, and performed a white balance test with a Dage camera. To control for variations in
stain color and contrast between different histology staining
sessions, one slide was chosen from each stain type (WG,
H&E and IP), and several fields of view were selected from
that slide. Images were taken with 10×, 20×, 40× and 100×
objectives for a total of six images per stain type.
Downloaded from jhc.sagepub.com by guest on March 26, 2015
235
Method for Consistent Images
Microscope Session Routine: All images were taken on
the same day. To simulate multiple sessions over more than
one day we did the following:
Automated Exposure and White Balancing
1. Camera was turned off intermittently between
images of different fields of view.
2. Light attenuation was set manually to factory recommended setting.
3. The camera was set to Auto Exposure and Auto
White Balance.
4. A non-tissue background area was found so that
Auto White Balance could correlate camera to color
temperature on a neutral area before finding the
specimen field.
5. Image was acquired.
Manual Exposure and White Balancing
1. Light attenuation was turned down (to lower the
voltage).
2. Auto Exposure and Auto White Balance were set
manually.
3. Light attenuation was turned up to factory recommended setting.
4. Camera was white balanced on a non-tissue area,
and a similar or identical field was found.
5. Camera exposure was set to a live histogram, with
shutter speed and gain alternately used to expand histogram to the setting with the greatest dynamic range.
Consistency in images can be measured by comparing a
high-quality image with other images of the same stain and
specimen using Delta E (DE), the unit of measurement in
the field of color science (Robertson 1990). This measurement indicates a change in color by comparing one image to
a target image (generally a high-quality image). The
Euclidean distances between similar colors when colors are
mapped to a three-dimensional space are compared between
the target image and comparison images. A DE of less than
1 U indicates no perceptual difference between images, and
a DE of 2 U or greater is commonly considered perceptually
significant. A change in color can be due not only to differences in hue but to changes in exposure and poor (or no)
white balancing.
For DE measurement, regions of interest (ROIs) in nonidentical images were outlined in Photoshop, such as red
blood cells, counterstained cells, non-tissue areas (background) and stained cells. ROIs were found via a macro
recorded in Photoshop, where tonal ranges instead of colors
identified ROIs to remove bias when attempting to choose
perceptually similar colors. When two or more images were
identical, the entire image was used instead of the ROI.
Photoshop Macro for Selecting Stained, Background, and
Cell Areas
1. Layers were created above the original image layer.
2. Using either the Black/White or Channel Mixer
functions, a grayscale image was created using color
components to create brighter or darker nuclei,
RBCs, or IP-stained areas.
3. Brighter/darker areas were selected through the use
of the Select Colors function with Highlights,
Midtones or Shadows from the drop down list.
4. Or, for background area values (white, non-tissue
areas), a threshold tool was used to select the brightest 5%–10% of pixels.
5. The consequent ROI selections were applied to the
original image layer for measurement.
6. Mean RGB values were found in the Histogram dialog box for the red, green and blue channels and
recorded in Excel, version 2003 (Microsoft
Corporation; Redmond, WA).
Stained images were both visually assessed and measured (for
DE values) to determine target images for each stain type.
Target images (images against which DEs were measured) created minimal variation of DE so that each stain type was measured from the same basis. RGB values were compared to RGB
values from the target image for each ROI from each stain. The
DE value was determined through RGB_DE_Calculation_
D50. DE values were recorded in Excel for each ROI. ROIs
from each feature were averaged to determine a final DE value.
Study 2: Post-Processing Comparison between Photoshop and
ChromaCal. A slightly underexposed and poorly white balanced WG-stained image taken at a microscope session was
chosen as a test slide for this comparison. During the same
microscope session, a calibration slide provided as part of
the ChromaCal system was also acquired (Fig. 1), to comply with required ChromaCal procedures.
After acquiring the calibration slide image, the steps
were as follows:
1. The image of the ChromaCal slide was opened in
ChromaCal software (version 1.4).
2. The grid was applied over the matrix of circles so
that software read color values.
3. The specimen image(s) were opened.
4. The images were saved individually, or automatically using the batch function.
5. The images were saved as TIFF files, allowing
ChromaCal software to apply a label on the images,
and automatically saved with a prefix to the filename to preserve the original.
ChromaCal software allows the user to proceed with postprocessing only when the grid is adequately placed over
Downloaded from jhc.sagepub.com by guest on March 26, 2015
236
Linden et al.
Figure 1. ChromaCal Calibration Slide with Enlarged Color
Matrix. The color circles of the matrix are composed of a
substance that effectively creates filters at incremental wavelengths
from deep blue to red. A column of grayscale circles (right)
incrementally vary transmission of light at company-determined
values similar to neutral density filters. Slide dimensions are similar
to conventional glass microscope slides. Color matrix shown
(180 microns) fits within image fields for 4× to 40× objectives; a
smaller matrix is available for greater magnifications. The image is
overexposed for the purpose of this figure to show darker colors.
color matrix, exposure is within the required range, and the
calibration slide image is within an acceptable linear regression (for red, green and blue channel measurements) to be
considered linear (Fig. 2).
Specimen images were post-processed either in
ChromaCal or in Photoshop using automated and semiautomated functions for color images with all functions at
default (factory) settings. Functions in Photoshop included
Auto Levels (Auto Tone in more recent versions of
Photoshop), Auto Color and an eyedropper tool method
(semi-automated), which required the selection of a pixel
that should be white on the image itself (so that the red,
green and blue levels for that pixel can be used to adjust
levels for the remaining pixels). Two user selections were
made for the Eyedropper Tool method.
Images were evaluated on the percent accuracy of background red, green and blue levels compared with neutral
(equal red, green and blue levels), the percent of the image
that is overexposed (contains saturated pixels), and the
mean maximum tonal level of background areas compared
to the ideal level at 240 for reproduction.
For background measurements, the Color Select tool
was used in Photoshop to select greater than 90% of the
background area. From selected background areas, red,
green and blue average values were found to determine
color shift as a result of a non-white balanced image, and
were also used to find the mean background values (from
the Histogram palette). Saturated pixels were found in the
Histogram palette (using Count). Tonal values were transcribed to Excel.
Study 3: Immuno-Peroxidase Density and Surface Area Measurements using Variable and Fixed Thresholds. A single IP-stained
paraffin section of bone marrow from our stained specimens was used. The image was acquired on an Olympus
BX51 with the Dage camera and a 20× oil objective lens. At
the same session, an image of the ChromaCal calibration
slide was acquired with a 10× lens.
This single image was post-processed in Photoshop to
generate typical imaging defects. The first session image
was not altered; the 2nd was brightened by +25 in Brightness/
Contrast function at Legacy setting (each pixel was made
25 pixel values brighter); the 3rd darkened by -40 in
Brightness/Contrast; the 4th adjusted in Curves function to
create a non-white balanced, red-shifted image (R=233,
G=226, B=238); the 5th made brown-shifted image (R=205,
G=186, B=167) and the 6th made a yellow-shifted image
(R=251, G=219, B=184).
One set of images were processed in ChromaCal software to create color corrected, white balanced and brightness matched images; the other set was left as is (with image
defects).
Images for Measurement. To obtain both morphometric
measurements of the IP stained areas in the images and to
find a mean density (in arbitrary pixel levels), non-stained
areas were converted to pure black color (mask) via a routine
in Photoshop. The mask was created as follows for images
with defects and ChromaCal corrected images (Fig. 3):
1. The original image was duplicated to a layer.
2. The duplicated image was blurred by a radius of 1 in
the Gaussian Blur dialog box.
3. A threshold was applied to the image in the
Threshold dialog box to create a binary (pure black
and pure white) image. The threshold was set one of
four ways:
i. Variably via manual settings to image with
defects (Photoshop).
ii. At a consistent level (128) for image with
defects (Photoshop).
iii. At a consistent level (132) for ChromaCalcorrected image.
iv. Variably via manual settings to ChromaCalcorrected image.
4. The image was inverted to create white IP areas on
a black background.
5. The layer was changed to the darken mode in the
Layers palette (a convolution algorithm to only
allow lighter values to appear from the original
image layer below).
6. The image was saved in the TIFF format for
analysis.
Images Measured in Fiji. Area and mean density measurements were made, using pixels2 as surface area units of
Downloaded from jhc.sagepub.com by guest on March 26, 2015
237
Method for Consistent Images
Figure 2. ChromaCal User Interface and Linearity Check Graph. The enlarged, lower left panel of this screen shot shows the quality
control score for matrix alignment, linearity and exposure. The graph on right is a representation of red, green and blue channel analysis
for determination of linearity. The graph can be activated in the ChromaCal Tools menu.
measurement for area, and pixel tonal values as measurements of relative brightness levels (brightness channel after
changing image type to hue, saturation and brightness channels).
Results
Study 1: Image Variability due to Imaging
System and Operator
Figure 3. Method for Creating a Mask. (A) The original image of
immunoperoxidase (IP)-stained bone marrow: anti-kappa light chain
antibody-stained cytoplasm of plasma cells. (B) Separation of IPstained areas from surrounding areas after using the Black and White
function in Photoshop. (C) Image after setting a threshold value. (D)
Mask image after application of Invert function on (C). (E) overlay of
mask (D) onto (A). (F) Area measured colored red. Scale, 50 µm.
A camera operator and microscopist who are familiar with
the camera software and the microscope can virtually eliminated user error. When the operator used a single, scientificgrade camera over several sessions using manual settings,
the average DE varied depending upon the stain: For WG
smears, the average DE was 7.3; for H&E, 5.1; and for IP,
3.4. A DE unit of less than one indicates no perceptual difference between images. In color science, a DE of two or
greater is commonly considered perceptually significant
(Ruppertsberg and Bloj 2007). A change in color can be due
not only to differences in hue, but to changes in exposure
and poor (or no) white balancing.
The variability of images far surpasses the expectation
that proper use of a camera and microscope will yield
results that do not require post-processing. Given that
expectation, the results should be images that are perceptually similar over more than one imaging session. Instead,
our study shows perceptually dissimilar images (Fig. 4).
The addition of over- and underexposure increases differences, as well as subtle, overall color shifts due to inconsistent white balancing.
When Auto Exposure and Auto White Balance settings
were used, we found that color inconsistencies and under
exposure were greater (Fig. 5). The average DE values rose
overall: for WG, the average DE was 7.7; for H&E, 10.3;
and for IP, 4.6. Furthermore, when more tissue filled the
Downloaded from jhc.sagepub.com by guest on March 26, 2015
238
Linden et al.
Figure 4. Image Variability Using Manual Camera Settings.
Variability shown for representative sessions (of 5 total sessions)
with Wright–Giemsa (WG)-stained bone marrow aspirates; two
Hematoxylin-Eosin (HE)-stained bone marrow trephine core
biopsies; and two sessions for immunoperoxidase (IP)-stained
bone marrow biopsies (CD61, stains megakaryocyte cytoplasm).
(A, B) Note color differences between (A) and (B) for WG, with
reduced RBCs red staining in (A) than in (B); background values
are also more greenish in (A) than in (B). (C–D) Between HEstained images, (D) shows more red overall. (E–F) Between IPstained images, (E) shows increased peach color over that in (F).
Scale (A, B), 50 µm; (C–F), 100 µm.
field (versus non-tissue background area), the images darkened overall. White balancing was also inconsistent,
depending upon the nature of the specimen and staining.
Study 2: Post-Processing Comparison between
Photoshop and ChromaCal
When images contain the defects described in this paper,
the defects are often corrected during post-processing.
Photoshop is widely used to post-process images.
Photoshop’s automated functions were compared to
ChromaCal’s automated functions for the correction of
Figure 5. Image Variability Using Automated Camera Settings.
Variability shown for representative sessions (of 5 total sessions)
with Wright–Giemsa (WG)-stained bone marrow aspirates; two
Hematoxylin-Eosin (HE)-stained bone marrow trephine core
biopsies; and two sessions for immunoperoxidase (IP)-stained
bone marrow biopsies (CD61, stains megakaryocyte cytoplasm).
(A, B) Between WG-stained images, (A) appears brighter than
(B), and has a greener background. (C, D) Between the HEstained images, the tissue area in (C) is darker than that in (D).
Between IP-stained images, (E) is brighter than (F). Scale (A, B),
50 µm; (C, D), 50 µm; (E, F), 100 µm.
image defects that include poor white balancing, saturated
pixels, and inconsistent brightness levels.
Average Red, Green and Blue channel pixel values in
background areas were measured to determine color shifts as
compared to neutral values to show efficacy of automated
white balancing. For the original image, the neutrality was
measured at 91% of a possible 100% neutral (Table 1). The
ChromaCal processed image is at 99%, and Photoshopprocessed images are as follows: Eyedropper Tool method
A, 100%; Eyedropper Tool method B, 99%; Auto Levels,
100%; and Auto Color, 99%.
To measure the retention of detail and slight gray values
in the whitest parts of the image, the percentage of pixels
Downloaded from jhc.sagepub.com by guest on March 26, 2015
239
Method for Consistent Images
Table 1. Image Data for Photoshop- and ChromaCal-Corrected Images.
Image Label in Figure 6
A
B
C
D
E
F
Images
Percent Neutral
Percent Saturated
Mean Brightness
Original
ChromaCal Processed
Photoshop Eyedropper Tool A
Photoshop Eyedropper Tool B
Photoshop Auto Levels
Photoshop Auto Color
91%
99%
100%
99%
100%
100%
3%
3%
30%
7%
2%
0%
213
234
255
253
248
243
Results for each image in Figure 6 for white balancing (percent neutral pixels), percent of saturated pixels at maximum level (pixel level = 255), and
mean brightness from background areas.
that saturated in any of the red, green or blue channels is at
3% for the original image. The ChromaCal processed image
is also at 3%, whereas the Photoshop-processed images are
as follows: Eyedropper Tool method A, 30%; Eyedropper
Tool method B, 7%; Auto Levels, 2%; and Auto Color, 0%.
The background mean brightness level was measured to
determine how closely the value matches the recommended
maximum value for any red, green or blue channel when
publishing manuscripts; the tonal value that is widely
accepted for printing press reproduction is at a maximum
level of 240 (Kelby 2003). The original image mean value
is slightly underexposed at 213. The ChromaCal processed
image measured 234, and Photoshop-processed images
measured as follows: Eyedropper Tool method A, 255;
Eyedropper Tool method B, 253; Auto Levels, 248; and
Auto Color, 243.
These results imply that Photoshop may not be a viable
means to post-process CB images (Fig. 6). If the nature of
correction is to white balance, correct unnatural colors, and
brighten images for optimal reproduction, this study shows
that Photoshop may not correct images as well as
ChromaCal. Only the Auto Color function in Photoshop
appeared to produce the most optimally corrected image
with virtually no saturated pixels, a brightness level close to
accepted value, and an optimally white balanced image.
However, the color shifted to unnatural colors most noticeably using this automated function (note, however, that in
working with other images using the Auto Color function, it
was determined that those images that had already been
white balanced may not shift in color), and colors were
more muted than in the ChromaCal-corrected image.
Study 3: Immuno-Peroxidase Density and
Surface Area Measurements using Variable and
Fixed Thresholds
To test the ChromaCal system, a single image was used and
then common user errors were applied to duplicate images,
as described in the Materials & Methods. From the inconsistent images due to common user errors, a set of
ChromaCal -processed images were created to evaluate the
software’s ability to make these images consistent (Fig. 7).
Figure 6. Image Variability due to Post-Processing. (A) Original
image (Wright–Giemsa-stained bone marrow aspirate smear). (B)
ChromaCal-processed image. Images (C–F) were post-processed
in Photoshop: (C) and (D) were created using the Eyedropper
Tool method, which sampled different pixels within background
areas of the image; the image in (E) was created using the Auto
Level function; and the image in (F) using the Auto Color function
for adjustments. Scale, 20 µm.
The need for consistent images is especially important
when measuring images for morphometry (area, length,
width, counts, etc.) and for optical density (Yagi and
Gilbertson 2005). For morphometry, differences in brightness level or color provide a means to separate stained
objects in a specimen image from adjacent non-stained
areas (Russ 2010). When brightness levels are used, a cutoff
level is set to include all tones that are either darker or
lighter (representing stained areas). The cutoff level is set
by the user in imaging software through the use of a threshold tool, or the cutoff is set through an automated method.
Downloaded from jhc.sagepub.com by guest on March 26, 2015
240
Linden et al.
Figure 7. Immunoperoxidase (IP) Density and Surface Area Measurements using Two Post-Processing Methods. Top series of images
(A1–A6) are typical inconsistencies found in Color Brightfield imaging when acquiring images over time. A1 represents the original image
(IP-stained bone marrow biopsy: anti-kappa light chain antibody-stained cytoplasm of plasma cells), with remaining images in that row
made variable in post-processing using Photoshop to simulate typical inconsistencies: A2 is brighter; A3, darker; A4, too red; A5, too
brown; and A6, too yellow. When setting thresholds manually (to create regions of interest from stained cells) from inconsistent images
(A1–A6), note that areas chosen (represented by red blobs) are not identical from (B1–B6), and measurements from these images vary
(blue lines on graphs). Once images are corrected using the ChromaCal method (C1–C6), and a single threshold is used to segment
the stained cells (represented by red blobs: D1–D6), measurements are nearly identical (red lines on graphs). Red blobs are a smaller
portion of the image (outlined area on A1 and C1) to aid in visual evaluation. Scale, 50 µm. Regions of interest (ROIs) from uncorrected
and corrected images are taken from lower right hand corner of images in rows A and C (depicted in A1 and C1). ROIs are presented
as magnified images in rows B and D.
When images are consistent, only a single threshold value
needs to be applied for all images; but, when images are
inconsistent, the user must determine a threshold for each
image (Haase et al. 1996), or the user must depend on the
accuracy of an automated thresholding function (Dowsett
et al. 2000; Brey et al. 2003; Seidal 2001). A single
Downloaded from jhc.sagepub.com by guest on March 26, 2015
241
Method for Consistent Images
threshold provides an objective means for setting a cutoff
level, as the manual method is a subjective method and thus
prone to error (Nobis and Hunsiker 2005).
Using both Photoshop and ChromaCal, we applied variable (setting by “eye”) and single thresholds to inconsistent
images and compared the results. As all inconsistent images
are derived from a single original image, the resulting mean
tonal density and surface area measurements should be the
same.
For mean tonal density readings, variable thresholds to
inconsistent images varied from the first reading by a maximum of -28.8% and a minimum of 6.3%. A single threshold
applied to inconsistent images varied from the first reading
by a maximum of -31.7% and a minimum of 4.2%. Variable
thresholds applied to ChromaCal-processed and consistent
images varied from the first reading by a maximum of
-3.0% and a minimum of 2.3%. ChromaCal-processed and
consistent images with a single threshold varied from the
first reading by a maximum of -2.3% and a minimum of
0.3%. The results illustrate nearly identical values from one
image to the next for ChromaCal-processed images, ranging from 150 to 155 on a 0–255 scale.
For surface area and morphometric readings, variable
thresholds to inconsistent images varied from the first reading by a maximum of -58.9% and a minimum of 12.6%. A
single threshold applied to inconsistent images varied by a
maximum of -67.5% and a minimum of 76.0%. Variable
thresholds applied to ChromaCal-processed and consistent
images varied from the first reading by a maximum of
-18.2% and a minimum of 29.9%. ChromaCal processed
and consistent images with a single threshold varied from
the first reading by a maximum of 0.3% and a minimum of
-0.3%.
For both morphometric and optical density measurements, the readings from the consistent images generated
by the ChromaCal system outperformed thresholds set
manually.
Discussion
Obtaining digital images in CB microscopy is an important
aspect of biomedical research and the clinical practice of
pathology. There have been remarkable changes in the field
of digital pathology, and it has improved the speed at which
digital images are acquired and shared (including online
and printed publications). Indeed, the ease of digital photograph and manipulation/editing has prompted many scientific journals to publish guidelines on the integrity and
management of digital images. Although the speed of image
acquisition has increased, and authors in the field take the
necessary steps to not intentionally manipulate their published images, there has been little effort to standardize
color, tonal levels, white balancing, and other critical components of digital CB photomicrographs (which is quite
dissimilar to that in the graphic design industry, where standardization is commonplace).
In the first study presented in this manuscript, we compared images from a single microscopist over multiple sessions. In most practices, one would assume that multiple
users would share systems. For the purposes of this study,
however, we elected to have a single microscopist obtain
the images for this portion of the project, to minimize interoperator variability. Regardless, even with an expert operator using a scientific-grade scope and camera, there was
distinct session-to-session image variability, independent of
whether the user manually adjusted the camera settings or
used automated settings for both exposure and white balance. This finding suggests that variability lies in the camera system, particularly when unable to obtain identical
exposures because of limitations in the camera software
(found with multiple camera systems in our experience).
Few research experiments can be completed in a single
microscope session and, for pathologists who read slides,
multiple sessions are the norm.
The second study presented compared two post-processing methods to evaluate image consistency. For this study, a
single, original (underexposed and poorly white balanced)
image of a Wright–Giemsa-stained bone marrow aspirate
was processed by two different methods: a widely available
raster graphics editor (Adobe Photoshop) and ChromaCal.
For the Photoshop-processed images, the original image
was adjusted by four different commonly used methods:
Eyedropper Tool method A, Eyedropper Tool method B,
Auto Levels, and Auto Color. Each of these methods
improved the original image variably, and the Auto Color
function appeared to produce the most optimally corrected
image with few saturated pixels, acceptable brightness, and
an optimally white balanced image. However, despite these
reasonable performance data, one could argue that the colors presented in the ChromaCal-adjusted image (Fig. 6B)
are much “truer” than those of the most technically optimal
Photoshop-processed image (Fig. 6F). It is noted, however,
that this is a small study comparing one image and two
methods.
For the final study, we demonstrated how post-processing can affect semi-automated quantification of immunoreactive cell density and surface area measurements. For this
study, a single image was intentionally adjusted to five different permutations to simulate the most typical imageto-image variability. Using ChromaCal to post-process, all
five permutations were corrected. Additionally when a single threshold was applied to all five permutations, there was
less variability using ChromaCal software versus the individual, manual threshold application in Photoshop.
Moreover, pixel levels and the surface area of immunoreactive cells measured were more consistent from image to
image using the ChromaCal system. Whereas this experiment represented a simulated microscopic session, the data
Downloaded from jhc.sagepub.com by guest on March 26, 2015
242
Linden et al.
demonstrate the utility of ChromaCal and how this could be
used to adjust tonal levels over more than one session to
approximate the same exposure and white balance. Similar
results were observed with morphometric measurements,
also demonstrating that post-processing in ChromaCal provided an objective method for setting a single threshold
value for images from more than one microscope session.
In designing the study, we intentionally chose to control
these variables to most fully demonstrate the problem of
inherent image-to-image variability, as well as the similarities and differences of the two post-processing methods
evaluated. Some of the limitations of this study inherent to
image acquisition include: 1) a single microscope/camera
tested, 2) a single expert microscopist, and 3) a small repertoire of images and stains.
In addition, we focused our analysis on the most common post-processing methods used by non-experts. More
sophisticated Photoshop functions would include color
matching (with the color match function), the use of
Photoshop’s companion program and its image correction
software (Camera Raw), and manual correction by adjusting individual color channels linearly.
We did not assess issues surrounding image integrity,
which would include methods to retain the original image,
and a means through which image corrections can be reported.
Both are included with the ChromaCal system without user
intervention, but were not described in the current study.
In an unexpected finding, approximately 50% of camera
systems we have used were configured (either with default
settings, or with user-initiated changes to software parameters, or otherwise) such that images captured were non-linear (data not published). Since no easy diagnostic method is
available (prior to ChromaCal), we suspect this non-linear
condition is often unknown to users of the imaging system
and those who post-process and evaluate the resulting
images. Generally, if a gamma setting is available in camera
software, adjusting the setting to a gamma of one may result
in linear images, depending upon the software manufacturer.
To measure linearity without the ChromaCal slide, users
could use methods that require more effort, such as taking
images of a step wedge with known optical density values,
and then comparing mean density values from images of
each step with known optical density values. When values
do not correlate, the camera system is not linear.
In summary, we present a new system for post-processing digital CB images. We are confident that the ChromaCal
method will make it easier for users to post-process digital
CB images in a consistent manner with improved color
standardization, preserved linear tonal levels, reproducible
brightness levels, and automated white balancing.
Acknowledgments
We would like to thank Huy Tran for his contributions to our
study.
Declaration of Conflicting Interests
The authors declared the following potential conflicts of interest
with respect to the research, authorship, and/or publication of this
article: The first and third authors confirm no conflict of interest.
Jerry Sedgewick works as a scientific consultant and provides
consulting services to Datacolor.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article, with the exception of payment for consulting (Jerry Sedgewick only).
References
Brey EM, Lalani Z, Johnston C, Wong M, McIntire LV, Duke PJ,
Patrick CW Jr (2003). Automated selection of DAB-labeled
tissue for immunohistochemical quantification. J Histochem
Cytochem 51:575-586.
Chieco P, Jonker A, Melchiorri C, Vanni G, Van Noorden CJ
(1994). A user’s guide for avoiding errors in absorbance
image cytometry: a review with original experimental observations. Histochem J 26:1-19.
Cromey DW (2010). Avoiding twisted pixels: ethical guidelines
for the appropriate use and manipulation of scientific digital
images. Sci Eng Ethics 16:639-67.
Dowsett M, Cooke T, Ellis I, Gullick WJ, Gusterson B, Mallon E,
Walker R (2000). Assessment of HER2 status in breast cancer: why, when and how? Eur J Cancer 36:170-176.
Dunning K, Safo AO (2012). The ultimate Wright-Giemsa stain:
60 years in the making. Biotech Histochem 86(2):69-75.
Foster B (1997). Optimizing light microscopy for biological and
clinical labs. Dubuque, Iowa: Kendall Hunt Pub Co.
Garcia JE, Dyer AG, Greentree AD, Spring G, Wilksch PA (2013).
Linearisation of RGB camera response for quantitative image
analysis of visible and UV photography: a comparison of two
techniques. PLoS ONE 8:1-10.
Haase AT, Henry K, Zupancic M, Sedgewick G, Faust RA, Melroe H,
Cavert W, Gebhard K, Staskus K, Zhang ZQ, Dailey PJ, Balfour
HH Jr, Erice A, Perelson AS (1996). Quantitative image analysis
of HIV-1 infection in lymphoid tissue. Science 274:885-1048.
Inoue S (1986). Video microscopy, New York: Springer.
Kelby S (2003). The Photoshop CS book for digital photographers, Thousand Oaks: New Riders.
Lee HC (2005). Introduction to color imaging science, New York:
Cambridge University Press.
Leong FJ, Brady M, McGee JO (2003). Correction of uneven illumination (vignetting) in digital microscopy images. J Clin
Pathol 56:619-621.
Luo MR, Clarke AA, Rhodes PA, Schappo A, Scrivener SAR, Tait
CJ (1991). Quantifying color appearance. Part 1. LUTCHI
colour appearance data. Color Res Appl 16:166-180.
Margulis D (2006). Professional photoshop: The classic guide to
color correction, 5th edition, Berkeley: Peachpit Press.
Martin C, Blatt M (2013). Manipulation and misconduct in the
handling of image data. Plant Cell 25:3147-3148.
Matsumoto B (2010). Practical digital photomicrography: photography through the microscope for the life sciences, Santa
Barbara: Rocky Nook.
Downloaded from jhc.sagepub.com by guest on March 26, 2015
243
Method for Consistent Images
Min H (2009). Modeling human color perception under extended
luminance levels. ACM Trans Graph 28:27.
Nobis M, Hunsiker U (2005). Automatic thresholding for hemispherical canopy-photographs based on edge detection. Agr
Forest Meteorol 128:243-250.
Price RL, Gray JW (2011). Basic confocal microscopy, New
York: Springer.
Robertson A (1990). Historical development of CIE recommended
color difference equations. Color Res Appl 15:167-170.
Rossner M, Yamada KM (2004). What’s in a picture? The temptation of image manipulation. J Cell Bio 166:11-15.
Ruppertsberg A, Bloj M (2007). Reflecting on a room of one
reflectance. J Vision 7:1-13.
Russ J (2010). The image processing handbook, Boca Raton: CRC
Press.
Sedgewick J (2008). Scientific imaging with Photoshop:
Methods, measurement and output, Berkeley: Peachpit
Press.
Viggiano JA (2004). Comparison of accuracy of different white
balancing algorithms as quantified by their color constancy.
Sensors and Camera Systems for Scientific, Industrial, and
Digital Photography Applications V: Proceedings of the SPIE
5301 Bellingham.
Yagi Y, Gilbertson JR (2005). Digital imaging in pathology: the case for standardization. J Telemed Telecare 11:
109-116.
Downloaded from jhc.sagepub.com by guest on March 26, 2015