Analysis of microscope images_FINAL PRESENTATION
-
Upload
george-livanos -
Category
Documents
-
view
29 -
download
1
Transcript of Analysis of microscope images_FINAL PRESENTATION
ANALYSIS OF MICROSCOPE IMAGES
George Livanos
Thesis presentationto
The Department of Electronics and Computer Engineeringin partial fulfillment of the requirements
for the degree of the Master Diploma
Technical University of Crete
Supervisory Committee:
Professor Zervakis Michalis (Supervisor) Professor Liavas Athanasios Professor Petrakis Euripides October 2013
Presentation scheme Introduction Thesis contribution Polarimetric imaging
Polarimetry modeling Preliminary results on tissue characterization
Microscope imaging Immunohistochemistry fundamentals Proposed cell segmentation and characterization Image clusteringWatershed transformActive contours membrane boundary estimationHer2 Grade evaluation
Results of tissue evaluation Segmentation performance Comparative estimations on Grade
Conclusion
Introduction Cancer is a disease that arises in molecular level In order to comprehend the cancer generation mechanism, a
deep study of the endocellular functions and extracellular interactions is necessary
Oncogene amplification and/or overexpression has been highly associated with breast tumor evolution
Medical advances have led to the identification of tumor biomarkers (onco-proteins) facilitating the understanding of the molecular basis of tumor progression and treatment response
Prognostic markers aim to objectively estimate the patient’s overall outcome
Predictive markers focus on the objective evaluation of the possible benefits from a specific clinical intervention
Introduction (cont’d) The HER2/neu oncogene is notable both for its role in the
pathogenesis of breast cancer and for its selection as a target of treatment
Overexpression of this receptor in breast cancer is associated with increased disease recurrence, poorer relapse-free survival and worse prognosis
Its predictive implications include resistance to hormonal therapy, resistance to chemotherapy , responsiveness to doxorubicin and, mainly, responsiveness to Trastuzumamb (Herceptin) therapies
Because of its prognostic role as well as its ability to predict response to Trastuzumamb, breast tumors are routinely checked for overexpression of HER2/neu
Introduction (cont’d) With the advent of molecular imaging, biological
and pathological disease processes can be determined in an intact living being by way of temporal and spatial distribution of exogenous probes or endogenous signals
Optical techniques span spatial scales from subcellular to organ level, relying on a disease-specific source of contrast as it modulates a measurable property of light
Backscattered optical polarimetric signatures from a target contain information related to target composition, structure, texture and geometry
Presentation scheme IntroductionThesis contribution
Thesis contribution In this thesis we study the process of tissue
characterization from both a macro and micro point of view
We study polarimetric imaging of tissue at macroscopic level and microscope imaging of stained cells at molecular level
By utilizing polarimetric imaging we extract characteristics of the tissue in a macroscopic level of analysis based on the optical identities of the tissue elements, studying their interaction with polarized light.
Information about the shape, size and physical state (solid or liquid) of such tissue elements can be extracted
By utilizing microscope imaging of tissue slides we are capable of studying tissue at a cellular level, obtaining information about its morphology, shape characteristics of cells and their distinct composite segments
Thesis contribution (cont’d) Diffused polarimetric reflection and backscattering provides
unique, discriminatory material signatures based on the depolarization of the impinging waves from different materials
By combining statistical analysis with polarimetric principles we develop an approach for analyzing the different properties of operational modalities and/or materials depicted in digital images
Qualitative and quantitative Her2 protein evaluation is also performed using immunohistochemistry (IHC) on frozen and archival tissues
Immunohistochemistry is a method for detecting specific antigens in tissues or cells and facilitates the identification of a large number of proteins, enzymes and tissue structures
Thesis contribution (cont’d)
The goal of this thesis is twofold
quantify the contrast effects and parameters of optical polarimetry for material characterization via statistical modeling and metrics and expand the outcome to biomedical applications
process IHC microscope images of breast tissues and automatically determine/define the impact of cancer on the specific organism using image processing techniques
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modeling
Polarimetric modeling Polarimetry is a technique that measures the extent to
which a substance interacts with plane polarized light It relies on the properties of polarization of backscattered
light and results in distinct signatures related to surface smoothness, orientation and target composition
Further contrast enhancement of the target can be achieved by modulating the background of the target through doping with polar and high-index-of-refraction molecules
When linearly polarized light is passed through a substance containing optically active molecules (chiral molecules) a rotation of the polarization vector take place. This phenomenon is called optical rotation or optical activity
Glucose molecules and most of the biological molecules such as proteins or enzymes are optically active molecules
Polarimetric modeling
LaserFilter
Polarizer, P1Retarder, R1Beam Expander
Retarder, R2
Polarizer, P2 CCD Camera
Phantom
633 nm laserbeam expander
aqueous optically active molecules
targetobjective lens
P1 R1
CCD
Stokes ParameterAnalysis
• A target phantom is stimulated by a laser beam and the backscattered light is captured through a “detection line”• the optical transceiver geometry consists of a transmitter generator and a receiver analyzer• the imaging system contains two arms; a polarization generating branch and a polarization analyzing branch•the system is utilized to image a specific ROI of the target, achieving optical performance of a microscope
Polarimetric modeling (cont’d)
n=index of refractionn=1.46, similar to hydrated collagen,L-Phenylalanine is an enzymen=1.55, similar to
calcified structures, salt and alcohol are molecular contrast agentsn1=1.54, similar to calcified structures n2=1.49, similar to hydrated collagenn3=1.65, similar to highly calcified mineralized structures
Emulate biological compounds: Micro calcifications depending upon their shape, geometry and composition can be classified as precursors of malignancies in breast mammography. Systematic differences between hydrated collagen in the intensities among the collagen of malignant, benign and normal tissue groups appear to be due to a significantly lower structural order within the malignant tissues
Concept
Polarimetric modeling (cont’d)The regions of different materials in the recorded image can be modeled by means of the intensity or the texture of the corresponding area
Materials that absorb different portions of the incident radiation would be recorded as image regions of quite different intensities
Apply modelling on the smoothed intensity
histogram (mean filtering)
Modalities that favor the dispersion and extensive diffuse of light at grain levels, then texture be preferred as a discriminant factor
Apply modelling on the
Variance of an image
Gaussian distribution as basis
Chi-square distribution as basis
Polarimetric modeling (cont’d)
22 2 (x-μ )(x-μ ) (x-μ ) 31 2f(x; A , , )= A exp(- ) + A exp(- ) A exp(- )1 2 3i 2 2 22σ 2σ 2σ1 2 3Target Background Remaining area
i i
2 2/2 1 /2 11 1/2 /22 2 21 2( ; , ) ( ) ( )1 2/2 /21 22 ( / 2) 2 ( / 2)1 2Target Background
n nf e en n
n n
A n
5x5 window convolution on mean and variance of the
original image
Estimation of distribution parameters
Description of image as a sum of basic
functions
Ai ->amplitude factors of distribution iμi ->mean of distribution iσ ι->variance of distribution i
Ai ->amplitude factors of distribution ini ->degrees of freedom of distribution i
No remaining area modeling
due to zero variance
Polarimetric modeling (cont’d)
The proximity of the two curves, as well as
the effectiveness of our approximation, is supported by the “goodness of fit
metrics”, the MSE and the SNR quantities, producing a value of order 5x10-3 for the MSE and an SNR db gain value of order
10db
First set of experiments
By decomposing the estimated
mixture models, we can easily
derive the image regions being
modeled by each individual
distributionInformation about the size, shape, texture of the object can be
extracted
Polarimetric modeling (cont’d)
The experimental results of this study emphasize the potential of the DOLP-polarimetric modality in biomedical applications, since the experimental setup and calibration of the structures can simulate the function and behavior of human body molecules and substances
Comparison of contrast for different concentrations of Phenylalanine using a 5x5 moving window
72,9268,7562,5058,3356,2549,00
100,0496,8395,6690,4889,3486,63
0,00
20,00
40,00
60,00
80,00
100,00
120,00
0,00% 2,22% 4,44% 6,66% 8,88% 11,10%
Concentration of Phenylalanine in aqueous solution (mg/ML)
Dist
ance
bet
wee
n th
e pe
aks
of
the
two
dist
ribut
ions
Use of mean
Use of variance
Increased contrast is
achieved by increasing the
concentration of optically active
molecules
Proceed to further experimentation on tissue samples
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Preliminary results on tissue characterization
Cancerous cells exhibit structural, biochemical and metabolic anomaliesPrecancerous lung epithelial tissues exhibit higher reflectance, increased DOP and larger retardance characteristics than healthy tissue, unlike other types of lung cancers of invasive nature. Different lung cancer types and subtypes appear distinct optical differencesLung tissue samples are utilized within the
experimental procedure, directly connecting Stokes parameter imaging to biomedicine
Preliminary results on tissue characterization (cont’d)
Characteristic difference of optical parameters between normal lung tissue and stage I carcinoma Introduce the potential of tissue discrimination through the proposed polarimetric imaging schemeThe key objective and hot prospect of our future study is to determine how lung anatomy and pathology relate to the optical parameters of lung cancer
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imagingImmunohistochemistry fundamentals
Immunohistochemistry fundamentals
GOAL
Detection of Her-2/neu protein in tissues Advanced image analysis techniques are adopted in order to accurately segment the cells within the sample IHC images and precisely extract their membrane contour and degree of staining
Main steps of an automated procedure for the assessment of HER2 status from a
representative Immunohistochemical
input imageThesis contribution
INNOVATION
Automated membrane contour evaluation without prior training of the algorithm on the image content Fusion and parameterization of advanced and fundamental image processing techniques
Immunohistochemistry fundamentals (cont’d) The effect of IHC testing on the tissue sample is to stain the
membranes of tumor cells, partially or completely The percentage of tumor cells that have completely or partially
stained membranes and the intensity of that staining determines the assigned score for HER2 overexpression according to a reference system
Score HER2 Status Staining Pattern
0 Negative: No staining observed, or membrane staining in less than 10% of tumor cells.
+1Negative: A faint/barely perceptible membrane staining detected in morethan 10% of tumor cells. The cells are only stained in part of the membrane.
+2Borderline: A weak to moderate complete membrane staining observed in more than 10% of tumor cells.
+3Positive: A strong complete membrane staining observed in more than10% of the tumor cells.
Immunohistochemistry fundamentals (cont’d)Score 0 Score
+1
Score +2
Score +3
SCORING
EXAMPLES
Immunohistochemistry fundamentals (cont’d)Original image in RGB model
IHC image with DAB staining protocol
Bad tissue preparation – color spread in neighboring areas
IHC image with PAP staining protocol
Bad tissue preparation – wrong cutting of tissue, intersecting nuclei
Tissue preparatio
n procedure affects the quality of the image
Immunohistochemistry fundamentals (cont’d)
Positive reactions with DAB (diaminobenzidine) are identified as a dark brown reaction product on the cell membrane Cell nuclei are identified as blue colored regions
Stained membrane
Cell nucleusThe evaluation procedure is usually performed qualitatively by a pathologist, who carefully observes the IHC samples via microscopy and manually calculates the presence of stained cells in the breast tissue
Intervariability of estimation according to the specialist’s
experience
In order to make immunohistoche
mical studies more objective,
quantitative techniques based
on computer-assisted
microscopy and image analysis
must be developed
IHC PROSPECT
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imagingImmunohistochemistry fundamentalsProposed cell segmentation and characterizati
on
Proposed cell segmentation and characterization
Segmentation scheme combining different image processing techniques
enhance the color difference between the membranes and
nuclei coloring
Improve Image
contrastExtract cell nuclei via mean-shift clustering
Extract cell boundaries
(membranes)via active contours initialized from
clustering results and watershed
transform
Apply mathematical morphology
and extract the final image
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imagingImmunohistochemistry fundamentalsProposed cell segmentation and characterizati
onImage clustering
Image clustering
The basic limitation of k-means clustering for a priori knowledge on the number of candidate classes is overcome utilizing the Mean-Shift clustering algorithm
The main idea is to treat the points as an empirical probability density function where dense regions in the feature space correspond to the local maxima or modes of the underlying distribution
Mean-shift clustering
Mean-shift clustering
Density Gradient Estimation
Data Density
modes
Image clustering (cont’d)Mean-shift clustering
1. Fix a window h around each data point2. Compute the mean of data within the window3. Shift the window to the mean and repeat till
convergence (no significant alteration in mean value is observed)
Repeat this procedure for each point. The number of different
mean values found represents the number of clusters
STEPS
The only parameter of the algorithm is the bandwidth h The number of classes is internally evaluated The basic drawback of the algorithm is the low speed of convergence The value of the bandwidth parameter is unspecified and application dependent
Remarks
Image clustering (cont’d)Mean-shift clustering
Given a kernel K, n data points {x1, x2,…, xn} and a bandwidth parameter h representing the window size, the Kernel density estimator for a given set of d-dimensional points is given as
2^
1 1
1 1( ) ( )n n
i id d
i i
x x x xf x K k
nh h nh h
Mean shift is assumed to lead to density maximization and can be based on gradient ascent on the density contour. Applying gradient to the kernel density estimator we obtain
2
2 2^ 1
21 1
1
mean shift vector
Assuming ( ) '( )
2 2( ) ( )
ni
in n ii iid d ni i i
i
g x k x
x xx g
hx x x xC Cf x x x g g xnh h nh h x x
gh
Image clustering (cont’d)Mean-shift clustering
21
^ ^2
, 2
2^1,
1
2
G
Setting G(x)=C g( x )2Gradient of f : ( ) ( ) ( )
2KDE of f with kernel G : ( )
Mean Shift Vector : m (x) =
G Gd
ni
G di
ii
Cf x f x m x
nh
x xCf x g
nh h
x xx g
h
1
2
1
^ ^2
G G, 2
2( ) ( ) m (x) m (x) proportional to gradient
n
i
ni
i
gd
xx x
gh
Cf x f x
nh
From x we move by the Mean shift vector towards the density maximum, i.e. cluster center
Mean shift is determined by the gradient
alone
objects in cluster 4
Image clustering (cont’d)
objects in LAB cluster 1 Original image objects in HSV cluster 1
objects in LAB cluster 2 Original image objects in HSV cluster 2
objects in LAB cluster 3 Original image objects in HSV cluster 3
Tissue sample Nucle
i
Intermediate areaMembran
e
Mean-shift clustering clustering
Five classes
produced, bandwidth h set to
0.15
Mean-shift segmentation of breast tissue image sample with Hue color
channel as input
objects in cluster 1objects in cluster 2
objects in cluster 3
objects in cluster 5
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imagingImmunohistochemistry fundamentalsProposed cell segmentation and characterizati
onImage clusteringWatershed transform
Watershed transform Segmentation via clustering produces broken boundaries at the points where color variety is met in the neighboring region Thus, a technique extracting closed boundaries is neededWatershed transform is a reference methodology regarding image segmentation The major idea beyond watershed transformation is based on the concept of topographic representation of image intensity Watershed segmentation displays more effectiveness and stability than other segmentation algorithms, producing closed contours and separating intersecting regions
Watershed transform (cont’d) Segmentation via clustering produces broken boundaries at the points where color variety is met in the neighboring region This limitation is overcome via the watershed transformWhen reporting on watershed transform, three fundamental notions arise; minima, catchment basins and watershed lines
Minima
www.mathworks.com
Watershed transform (cont’d)
(1) minima, points belonging to the different minima(2) catchment basins, points at which water would certainly fall
in a single minimum(3) watershed lines, the highest intensity level points at which
water would have equal probability to fall in more than one minimum, i.e. the separating boundaries of segmented basins
Intensity image viewed as a 3D depth/basin structure
The catchment basin CB(mi) of a minimum mi is defined as the set of points which are topographically closer to mi than to any other regional minimum mj:
The watershed of f is the set of points which do not belong to any catchment basin:
i f j f( ) j I\{i}|: f(m )+T ( , ) f(m )+T ( , )i i jCB m x m x m
( ) ( ) ii I
Wshed f D CB m
Watershed transform (cont’d)Approaches for image
segmentation
Start locating the basins and then
find the watershed lines as a set
complement of them
Partition the image into basins
and extract watersheds by
boundary detection
An appropriate labeling of the resulting regions is necessary All points belonging to a given catchment basin have the same unique labelPoints of different catchment basins have different labelsSometimes the watershed transform is not applied directly to the original image, but to its (morphological) gradient producing watersheds at the points of grey value discontinuity
Watershed transform (cont’d)
Key sub-regions or single points within each basin that significantly differ from the remaining basin region
Numerous local minima in the input image, thus regions are divided to smaller and smaller regions
Marked-watershed transform
Watershed transform
Markers
Marker extraction methodsH-minima transform: suppresses all minima in the intensity image whose height is less than a threshold h H-domes transform: finds local maxima using the morphological reconstruction operator
Watershed transform (cont’d)H-domes transform Difference of Blue channel from its
maximum
Input for the watershed transform Watershed lines
Markers for cell nuclei
Input image for the watershed transform
Differece image for the watershed transform
Contours extracted via the wateshed transform
Presentation scheme IntroductionThesis contributionPolarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imagingImmunohistochemistry fundamentalsProposed cell segmentation and characterizatio
nImage clusteringWatershed transformActive contours membrane boundary estimation
Active contours membrane boundary estimationActive contours
modelContour-based techniques are well established in international bibliography, providing accurate and robust results even in noisy environmentHowever they suffer from initialization, local minima and stopping criteria problemsGlobal minimum energy-searching methods have been proved effective in overcoming local minima problems, leading to robust convergence regarding the final contour extractiona curve C is represented via a function φ (the level-set
function) as C={(x, y)|φ(x, y)=0}
Consideration
(x, y) are coordinates in the image plane Curve representation in terms of function φ()
the evolution of the curve is given by the zero level curve at time t of function φ(x, y, t)φ <0 denotes points outside the curve φ >0 represents points belonging to the internal area of the curve
Active contours membrane boundary estimation (cont’d)
At any given time, the level set function simultaneously defines an edge contour (φ=0) and a segment of the image (φ≠0)The curve is evolving according to the partial differential equation below, iteratively converging to a meaningful segmentation of the image.
0, φ(x,y,0)=φ ( , )F x yt
Assuming that the image I consists of two regions of approximately constant distinct intensities I1 and I2 and that the object of interest is represented by the region of value I1 (inside curve C), the “fitting energy functional” is denoted
where F denotes the speed of the curve evolution
2 2
1 2 2 1 1 2 2( , , ) ( ) ( , ) ( , )
( ) ( )insideC outsideC
F c c C F C I x y c dxdy I x y c dxdy
Length C v Area inside C
where C is any variable curve except for the object boundary C0, constants c1 , c2, are the averages of I inside and outside C respectively and μ,ν,λ1,λ2 are non-negative fixed user defined parametersThe contour of the foreground region is the solution of the minimization problem inf C (F(c1, c2, C)) ≈0≈F(c1, c2, C0)
Active contours membrane boundary estimation(cont’d)
Color Pre-processing and Transformation to
Improve ContrastClustering for Region-
Based Segmentatio
n
Edge-Based Segmentation
for contour Refinement
Active contours membrane boundary estimation(cont’d)Active contours membrane boundary estimation
Active contours approaches (snakes) suffer from initialization , especially in the cases of unknown number of objects within the image Clever and targeted initialization by fixed-sized circle regions in the centre of each estimated cell region These circle regions originate from the application of mean-shift clustering and the refinement of the areas using mathematical morphology Too big or too small clustered segments are rejected as they are impossible of representing a cell nuclei area The colour information derived from the A and B colour bands of the equalized immunohistochemical image along with the initialized contour constitutes the input of the active contours segmentation algorithm A and B channels marked with watershed transform in order to secure convergence
Input Image initial contour
200 400 600
100
200
300
400
969 Iterations Global Region-Based Segmentation
Watershed marked equalized image in the
LAB space
Initial contours derived from cell nuclei detected
after clusteringInput for
active contours model
+Input marked image for active contours algorithm
Active contours membrane boundary estimation(cont’d)Active contours
segmentation Having completed the pre-processing steps, apply the active contours segmentation in order to extract the cell areas During the segmentation algorithm, the initial curve evolves until it terminates to closed boundaries forming the membranes of candidate cells Convergence is achieved until no significant change in the size of the evolved curve takes place
Input Image initial contour
200 400 600
100
200
300
400
969 Iterations Global Region-Based Segmentation
Input Image initial contour
200 400 600
100
200
300
400
969 Iterations Global Region-Based SegmentationConverged estimation of φ
superimposed on the original image
Internal area of φ (φ>0)
Active contours membrane boundary estimation(cont’d)Active contours segmentation – application
of morphologyApplication of mathematical morphology based on the characteristics of the extracted regions is performed Irregular shaped and sized regions are rejected, refining the segmentation result and improving its accuracy Regions with no internal blue regions (cells without nuclei) are also rejected
Extracted quantities
Labelled cell regions having nuclei Contours of the segments forming the cell membrane Number of pixels belonging to the cell boundary The intensity value of each pixel of the boundary Area and centroids of each labelled regions
Active contours membrane boundary estimation(cont’d)
Finally selected components after rejecting non candidate cells due to their shape characteristics
Segmented cell components
Finally extracted components
Finally marked candidate components after rejecting non candidate cells due to their shape characteristics and marking nuclei Initially segmented candidate cells Marked candidate nuclei
Candidate cell segments having nuclei
Finally estimated cell segments
Active contours membrane boundary estimation(cont’d)
Motivation and final result of the segmentation scheme
Better discriminate cell and nuclei areas based on color difference
Correct image defects
Extract closed boundaries,
adaptive to image content
Filter out irrelevant areas, yet improving
segmentation accuracy
LAB color model
selection
Final output image applying the watershed transform
Finally calculated cell border superimposed on the
original image
Histogram equalization
Mean-shift clustering
Intermediate step, used for
active contours
initialization
Active contours
model
Mathematical
morphology
Despite the increased level of region intersection, the
ambiguity of area borders is reduced at the perimeter of
each cell
Watershed transformati
on
Intermediate step, feed snakes with boundary estimation
Presentation scheme Introduction Thesis contribution Polarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imaging Immunohistochemistry fundamentalsProposed cell segmentation and characterizationImage clusteringWatershed transformActive contours membrane boundary estimationHer2 Grade evaluation
HER2 Grade Evaluation Membrane staining evaluation and IHC score
assignment
Final output image applying the watershed transform
Segmentation result
Extracted quantities
Labelled cell regions having nuclei Contours of the segments forming the cell membrane Number of pixels belonging to the cell boundary The intensity value of each pixel of the boundary Area and centroids of each labelled regions
+ =Number of cells forming the tissue Estimation of staining from intensity thresholding Estimation of staining completeness Number of complete stained cells
Her2 Grade evaluation
Presentation scheme Introduction Thesis contribution Polarimetric imagingPolarimetry modelingPreliminary results on tissue characterization
Microscope imaging Immunohistochemistry fundamentalsProposed cell segmentation and characterizationImage clusteringWatershed transformActive contours membrane boundary estimationHer2 Grade evaluation
Results of tissue evaluation Segmentation performance
Segmentation performance
Sample
1
Sample
3
Sample
5
Sample
7
Sample
9
Sample
11
Sample
13
Sample
15
Sample
17
Sample
19
Sample
21
Sample
23
Sample
25
Sample
27
Sample
29
Sample
31
Sample
33
Sample
35
Sample
37
Sample
390.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
Precision of cell segmentation procedure
Sample images (Sample number)
Prec
isio
n (%
)
TP = correctly detected regions as cellsFP = regions that have been wrongly detected as cellsPrecision=TP/(TP+FP)=Correctly detected cells / total regions detected
Possibility of
detecting a cell areaGiven the image content complexity,
this is a very encouraging result!
Average value=54,22%
Segmentation performance (cont’d)
Sample
1
Sample
3
Sample
5
Sample
7
Sample
9
Sample
11
Sample
13
Sample
15
Sample
17
Sample
19
Sample
21
Sample
23
Sample
25
Sample
27
Sample
29
Sample
31
Sample
33
Sample
35
Sample
37
Sample
390.00%
10.00%20.00%30.00%40.00%50.00%60.00%70.00%80.00%90.00%
100.00%
True positive rate (Recall) of cell segmentation pro-cedure
Sample images (Sample number)
Reca
ll (%
)
TP = correctly detected regions as cellsFN = regions forming cells that have not been detectedRecall =TP/(TP+FP)=Correctly detected cells / true number of cells
Hit rate of cell
detection
The algorithm reveals a quite large fraction of relevant results retrieved ->goal of cell segmentation
Average value=84,62%
Segmentation performance (cont’d)
Sample
1
Sample
3
Sample
5
Sample
7
Sample
9
Sample
11
Sample
13
Sample
15
Sample
17
Sample
19
Sample
21
Sample
23
Sample
25
Sample
27
Sample
29
Sample
31
Sample
33
Sample
35
Sample
37
Sample
390.00%
10.00%20.00%30.00%40.00%50.00%60.00%70.00%80.00%90.00%
100.00%
F-measure of cell segmentation procedure
Sample images (Sample number)
F-m
easu
re (
%)
F-measure=2*Precision*Recall/(Precision +Recall) An estimat
e of detectio
n accurac
y
The balanced test statistic confirms the encouraging segmentation results
Average value=65,92%
Segmentation performance (cont’d)
There is no numerical gold standard in cell segmentation performance estimation The vast majority of the techniques focus on the visual comparison between the extracted and the estimated result A medical expert’s evaluation is considered as a ground truth
Intervariability between specialists’ outcomes, subjective
estimation
Segmentation performance (cont’d)
However, we compare our methodology to two representative techniques
A specialized one
A classic oneFocus on IHC membrane
evaluation through color deconvolution,
skeletons and particle filtering
Marker-controlled watershed applied on
the gradient of the image marking the foreground and the
backgroundGroun
d truth
The specialist initially marks all candidate cell regions and then selects the true cell areas. Thus, the True Negative notion can now be
adopted.Two images are selected from the dataset: a ”mean” image
corresponding to relatively less complex content and a “difficult” one, where irregular regions mislead our algorithm
Segmentation performance (cont’d)Manually segmented
image
Color deconvolution segmentation
Marker-controlled watershed
Output of our methodology
Output image applying the marked watershed transform
MEAN
IMAGE
Segmentation performance (cont’d)Manually segmented
image
Color deconvolution segmentation
Marked watershed segmentation
Output of our methodology
Final output image applying the watershed transform
DIFFICULT
IMAGE
FP regions
Segmentation performance (cont’d)
Quantitative comparison of our methodology to other representative techniques
Precision Recall Specificity Accuracy FP rate F measure
Mean case
Our methodology74,56% 93,41% 80,54% 85,42% 19,46% 82,93%
Color deconvolution
63,97% 95,60% 67,11% 77,92% 32,89% 76,65%
Marked watershed38,39% 47,25% 53,69% 51,25% 46,31% 42,36%
Difficultcase
Our methodology69,37% 82,80% 53,42% 69,88% 46,58% 75,49%
Color deconvolution
66,39% 84,95% 45,21% 67,47% 54,79% 74,53%
Marked watershed20,31% 14,29% 32,00% 22,29% 69,86% 16,77%
TP = correctly detected regions as cellsFN = regions forming cells that have not been detected TN = regions that the specialist has finally rejected as cellsFP = regions that have been wrongly detected as cells
Our methodology is less affected by misleading color
areas!
Presentation scheme Introduction Thesis contribution Polarimetric imaging
Polarimetry modeling Preliminary results on tissue characterization
Microscope imaging Immunohistochemistry fundamentals Proposed cell segmentation and characterizationImage clusteringWatershed transformActive contours membrane boundary estimationHer2 Grade evaluation
Results of tissue evaluation Segmentation performance Comparative estimations on Grade
Comparative estimation on Grade
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%0.00%
20.00%40.00%60.00%80.00%
100.00%
54.29%
74.29%82.86%
91.43% 91.43%97.14%100.00%
ROC Curve for IHC classification methodology
1,00%8,00%10,00%12,00%14,00%16,00%20,00%
False Positive Rate (1-Specificity)True
Pos
itiv
e Ra
te (S
enso
tiv-
ity)
Score thresho
ld
Performance metrics of the implemented classification procedure and the reference methodology
Method Precision Recall FP rate Accuracy Specificity F-measure
Our method 96,97% 91,43% 20,00% 90% 80,00% 80,00%
Reference method 87,5% 100% 100% 87,5% 0,00% 93,33%
Promising classifier Overestimates positive cases
Operating point
Presentation scheme Introduction Thesis contribution Polarimetric imaging
Polarimetry modeling Preliminary results on tissue characterization
Microscope imaging Immunohistochemistry fundamentals Proposed cell segmentation and characterization Image clusteringWatershed transformActive contours membrane boundary estimationHer2 Grade evaluation
Results of tissue evaluation Segmentation performance Comparative estimations on Grade
Conclusion
Conclusion We propose a novel attempt for breast
cancer cell-nuclei segmentation and membrane border extraction
The innovative part of the methodology is on fusing information from region and contour-based techniques on different color spaces
The performance of the contour based approach is increased by the proposed initialization scheme
Preliminary results from the analysis of immunohistochemical images confirm the prospect of the methodology
Conclusion cont’d In algorithmic terms, different
segmentation approaches exploit different information and have different effects
Edge-based approaches (active contours) can resolve boundary uncertainties and derive well defined membrane edges.
Need fusion of these approaches Need good pre-processing in terms of color
and intensity enhancement IHC microscope imaging needs careful and
extensive analysis due to complexity of image
Conclusion Exploitation of several pieces of
information in a modular form enables the progressive derivation and refinement of results
Image structure is obtained from intensity
Cell nucleus and membrane appear in different colors in large distance on the color space.
The cooperation of image processing with pathology specialists can give rise to key findings in cancer diagnosis and treatment
The encouraging results of the proposed scheme give rise for its improvement and incorporation to laboratory assessments
PublicationsConference and journal publications referring to the presented work:Journals
IOP PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY Meas. Sci. Technol. 20 (2009)(12pp) Impact factor 1.317 "Stokes parameter imaging of multi-index of refraction biological phantoms utilizing optically active molecular contrast agents“
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 11, 2010 Impact factor 1.106 "Efficient Molecular Imaging Techniques Using Optically Active Molecules"
MEASUREMENT SCIENCE AND TECHNOLOGY Meas. Sci. Technol. 22 (2011) 114018 (12pp) Impact factor 1.494 "Polarimetric phenomenology of photons with lung cancer tissue”
International Society for Analytical Cytology, Cytometry Part A 71A:439–450 (2007) Impact factor 0.939 " Automated Analysis of FISH and Immunohistochemistry Images: A Review "
Image Processing, IET, August 2011, Volume: 5, Issue: 5, page(s): 429 – 439 Impact factor 0.639 " Modeling the characteristics of material distributions in polarimetric images "
PublicationsConference and journal publications referring to the presented work:
Conferences IST 2010, IEEE International Conference, Thessaloniki, 2010 " Backscattered polarimetric detection from biological tissue“
www.springerlink.com : 5th European IFMBE Conference, IFMBE Proceedings 37, pp. 381–384, 2011 "Histogram modeling of polarimetric images for analysis of material properties“
IST 2011, IEEE International Conference, Batu Ferringhi, Penang, Malaysia,2011 "Near Infrared Light Interaction with Lung Cancer Cells"
Thank you!