hourglass shapes in rank grey-level hit-or-miss …...hourglass shapes in rank grey-level...
TRANSCRIPT
Hourglass shapes in rank grey-level
hit-or-miss transform for membrane
segmentation in HER2/neu images
Marek Wdowiak, Tomasz Markiewicz, Stanislaw Osowski, Zaneta Swiderska,
Janusz Patera, Wojciech Kozlowski
Warsaw University of Technology, Military Institute of Medicine, Military University of Technology, Warsaw, Poland
• Segmentation of the thin, non-continuous and highly
variable objects is a difficult task in mathematical
morphology application.
• To such problems belongs the membrane segmentation of
the cells in histopathology Human Epidermal Growth
Factor Receptor 2 (HER2/neu) images related to the breast
cancer.
• The histopathological evaluation of such
immunochemistry stains specimens is the most common
task for pathologists.
Introduction
| Introduction | Methods | Results | Conclusion | Acknowledgement |
Introduction
Four categories in grade scale are recognized:
• 0 (no membrane staining is observed or membrane staining is
observed in less than 10% of the tumor cells),
• 1+ (a barely perceptible membrane staining is detected in more
than 10% of tumor cells, the cells exhibit incomplete membrane
staining),
• 2+ (a weak to moderate complete membrane staining is observed
in more than 10% of tumor cells),
• 3+ (a strong complete membrane staining is observed in more
than 10% of tumor cells).
| Introduction | Methods | Results | Conclusion | Acknowledgement |
Introduction
As shown, the categorization
criteria are very subjective and
may lead to the significant
differences in assessment of
the particular cases. Especially
the distinction between weak
and strong, and the continuity
of a membrane staining may
result in a large variation in
their interpretation.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
Methods – the proposed algorithm of image processing
The image processing pathway was composed on:
a) enhancement of an image contrast,
b) creating a set of different colour space representations and
classification of image pixels into the nuclei, membrane
with positive reaction, and the remaining regions,
c) segmentation of the cell nuclei,
d) recognition of the stained cell membranes based on the
modified rank grey level hit-or-miss transform with the
new proposed shape patterns,
e) allocation of the parts of membranes to the separate cells.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – enhancement and colour representations
• The contrast enhance was realized by applying an automatically
computed contrast stretch and normalizing the colour components
to fulfill their ranges.
• Next, the nine sub-image samples were prepared and manually
segmented into the blue nuclei (class 1), reactive brown
membrane (class 2), and the remaining areas (class 3) to establish
the most adequate pixel colour components for a classifier.
• The following colour spaces were taken into account: RGB,
CMYK, HSV, YCbCr, CIE Lab, CIE Lch, CIE uvL, CIE XYZ.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – enhancement and colour representations
• Their abilities to differentiate the specific classes were evaluated
using an area (AUC) under Receiver Operating Characteristics
(ROC) curve .
• Also, the cross-correlation between the candidate features was
studied in order to select not only the best ones, but also with the
least correlation between themselves.
• Finally, a set of six features composed of two the best ones for
each class was found and used as a feature vector for a classifier.
• The classification was performed using the SVM with a Gaussian
kernel function.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – learning data
The original images and
manually marked classes
comes from the nine
samples of HER2/neu
image as the learning
data.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
• The binary nuclei mask obtained from the
classifier is used for the cell nuclei segmentation.
• The morphological closing operation was
applied to reduce a noise.
• For separation of the connected nuclei the
distance map was build and each extended
regional maxima with the selected h value was
recognized as the centre of cell nucleus.
• Finally, the restriction to a single cell nucleus
area was applied to eliminate the non-cancer cell
masks.
Methods - segmentation of cell nuclei
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods - intensity map for membrane segmentation
• The crucial input signal to this step
can be an intensity map restricted
comes from a single colour channel,
e.g. luminance or yellow
component, as well as composed
from a set of colour channels.
• The presented sample intensity map
is the element-wise product of Y
channel from CMYK and u channel
form CIE uvL colour space.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – novel hourglass shapes for discovering the line structures
• To discover the line structures the appropriate
composed structuring elements should be
applied.
• Due to a high variability of a cell membrane
geometry, a flexible shapes are necessary.
• We proposed and defined the novel hourglass
shapes as they are presented on the figures.
• The central region (marked on red) is treated as a
foreground (FG), the hourglass shape is marked
in grey (HG), and the white areas represent
background (BG).
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – novel hourglass shapes for discovering the line structures
• The second set of shapes were the
half-hourglass structures presented
on the right.
• When the two hourglass shapes
oriented horizontally and vertically
covered the most cell membrane
orientations, the half-hourglass are
in four directions and each of them
required a highest correspondence
of membrane segment shape with
them.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – grey-level hit-or-miss transforms – related works
In the literature the most closely to the examined problem are:
• Blur grey-level hit-or-miss transform (Bouraoui et al., 2010) ,
defined as:
𝜂[𝑉,𝑊;𝐺,𝐻]𝑠 𝐼 𝑝 =
𝐼 ⊕ 𝐺 ⊖ 𝐴 𝑝 − 𝑎 𝑖𝑓 𝐼 ⊕ 𝐺 ⊖ 𝐴 𝑝 − 𝑎
≥ 𝐼 ⊖ 𝐻∨ ⊕ 𝐵∨ 𝑝 − 𝑏
⊥ 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where the dilation eliminates dark noise in the bright FG, while the
erosion 𝐼 ⊖ 𝐻∨ eliminates bright noise in the dark BG.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – grey-level hit-or-miss transforms – related works
The other extension of HMT is the rank (gray-level) hit-or-miss
(Soille, 2003):
[𝑈𝐻𝑀𝑇 𝑓 ] 𝑥 =
𝜁𝐵𝐹𝐺𝑘𝐹𝐺 𝑓 𝑥 − 𝜁𝐵𝐵𝐺𝑘𝐹𝐺′ 𝑓 𝑥 𝑖𝑓 𝜁𝐵𝐵𝐺𝑘𝐹𝐺
′ 𝑓 𝑥
< 𝜁𝐵𝐹𝐺𝑘𝐹𝐺 𝑓 𝑥
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where the rank vector determines the uncompleted inclusion of
FG/BG in the spatial location on the image.
This formulation is the most closely to requirement of our problem.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – grey-level hit-or-miss transforms – our modification
In our study we propose a slight modification to the Soilles’ UHMT,
that can be defined in the form:
where the result is in a binary form directly defined on the grey-
level by the h value. Instead of a rank filter, for the FG region we
applied the average value as a reference level.
Finally, a watershed algorithm is performed to established the cell
contours and allocation of membranes in the separate cells.
hXXHMT FGkBhk BGBGBG ,,,B
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Methods – the proposed scheme of the algorithm
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Scheme of the Algorithm | Enhancement | Learning Data | Cell Segmentation | Membrane Segmentation | Hourglass Shapes | Hit-or-Miss |
Results – material
• The materials used in experiments come from the archive of the
Pathomorphology Department in the Military Institute of Medicine in
Warsaw, Poland.
• Twenty cases of the breast cancer of HER2/neu preparations without
any artefacts representing 1+, 2+, and 3+ grades were selected. The
paraffin embedded breast tissue were stained in a standard way
according to the Ventana PATHWAY anti-HER-2/neu (4B5) Rabbit
Monoclonal Primary Antibody protocol.
• The specimen images were registered on the Olympus BX-61
microscope with the DP-72 colour camera under the magnification
400x and resolution 1024x768 pixels.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Results – feature selection for the classification process
On the basis of the
presented results the
following features have
been selected: R and B
components from RGB
space, u and L
components from uvL
representation, b
component from CIE
Lab and Y component
from CMYK.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Results – hourglass shapes for HMT
The hourglass shapes
were adjusted to: FG
size 3×3 pixels, the
distance between FG
and BG equal 4
pixels, mask size
21×21 pixels.
The rank value was
equivalent of 80 %,
the normalized h
value was 0.2
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Results – half-hourglass shapes for HMT
The half-hourglass
shapes were
adjusted to: FG size
3×3 pixels, the
distance between
FG and BG equal 4
pixels, mask size
21×21 pixels.
The rank value was
equivalent of 80 %,
the normalized h
value was 0.2 | Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Results – cell segmentation and allocation of membranes in separate cells
• Next, based on the
watershed, the cell
outlines were found.
• Finally, the recognized
immunoreactive
membrane segments
were allocated to the
specific cells.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Graphical Results
Expert Algorithm Expert Algorithm Expert Algorithm
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Graphical Results
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Numeric Results
The numerical
results of
membrane
continuity
estimation made by
expert and our
system at
application of
hourglass and half-
hourglass shapes.
Case Continuity
Expert’s result Hourglass Error(%) Half-hourglass Error(%) 1 0,404 0,383 -0,020 0,351 -0,053
2 0,316 0,367 0,051 0,335 0,019
3 0,293 0,423 0,131 0,388 0,096
4 0,359 0,534 0,175 0,474 0,116
5 0,316 0,371 0,055 0,332 0,016
6 0,304 0,446 0,142 0,401 0,097
7 0,322 0,330 0,008 0,289 -0,033
8 0,353 0,378 0,025 0,339 -0,014
9 0,232 0,309 0,077 0,279 0,047
10 0,328 0,450 0,122 0,409 0,080
11 0,308 0,380 0,072 0,343 0,036
12 0,191 0,295 0,104 0,271 0,080
13 0,256 0,339 0,083 0,314 0,058
14 0,298 0,431 0,134 0,396 0,098
15 0,373 0,490 0,116 0,447 0,074
Mean absolute error 8.8% 6.1%
Standard deviation 4.8% 3.2%
| Introduction | Methods | Results | Conclusion | Acknowledgement |
| Material | Feature Selection | Hourglass Shape Application | Cell Segmentation | Graphical Representation | Numeric Results |
Conclusion
• The new approach to the automatic evaluation of the HER2/neu
membrane staining in the breast samples was proposed.
• The important point in this approach is the application of the
hourglass-shape structuring elements in the rank grey-level hit-
or-miss transform for the analysis of the image.
• The experimental results have shown high efficiency of image
segmentation with respect to the nuclei and membrane
localizations.
| Introduction | Methods | Results | Conclusion | Acknowledgement |
Acknowledgement
This work is supported by the National Centre for Research and Development (Poland) by the grant PBS2/A9/21/2013 in the years 2013-2016. Thank you very much for your attention.
| Introduction | Methods | Results | Conclusion | Acknowledgement |