masters' whole work(big back-u_pslide)
TRANSCRIPT
Computer Assisted Screening of Microcalcifications In Digitized Mammogram For Early Detection of Breast CancerThesis Presentation
Nashid AlamRegistration No: [email protected]
Supervisor: Prof. Dr. Mohammed Jahirul Islam
Department of Computer Science and EngineeringShahjalal University of Science and TechnologyFriday, December 25, 2015
Driving research for better breast cancer treatment “The best protection is early detection”
010
2030
020
40-0.1
0
0.1
5; 0.5; 0.7854
5; 0.5; 0.7854
Introduction
Breast cancer:The most devastating and deadly diseases for women.
o Computer aided detection (CADe) o Computer aided diagnosis (CADx) systems
We will emphasis on :
Background Interest
Background Interest
Interest comes from two primary backgrounds
1. Improvement of pictorial information- - For Human Perception
How can an image/video be made more aesthetically pleasing
How can an image/video be enhanced to facilitate:extraction of useful information
Background Interest
Interest comes from two primary backgrounds
2. Processing of data for:Autonomous machine perception- Machine Vision
Micro-calcification
Mammography
Mammogram
Micro-calcification
Background knowledge
Micro-calcification
Micro-calcifications :- Tiny deposits of calcium- May be benign or malignant- A first cue of cancer.
Position:1. Can be scattered throughout the mammary gland, or 2. Occur in clusters.(diameters from some µm up to approximately 200 µm.)3. Considered regions of high frequency.
Micro-calcification
They are caused by a number of reasons:
1. Aging –The majority of diagnoses are made in women over 50
2. Genetic –Involving the BRCA1 (breast cancer 1, early onset) and
BRCA2 (breast cancer 2, early onset) genes
Micro-calcifications Pattern Determines :The future course of the action-
I. Whether it be further investigatory techniques (as part of the triple assessment), or
II. More regular screening
Mammography
Background knowledge
Mammography Machine
Mammography
USE:I. Viewing x-ray imageII. Manipulate X-ray image on a computer screen
Mammography :
Process of using low-energyx-rays to examine the human breast
Used as a diagnostic and a screening tool.
The goal of mammography :The early detection of breast cancer
Mammography Machine
Mammogram
Background knowledge
mdb226.jpg
Mammogram
Mammogram:An x-ray picture of the breast
Use:To look for changes that are not normal.
Result Archive:The results are recorded:
1. On x-ray film or 2.Directly into a computer
mdb226.jpg
Literature Review
To detect micro-calcifications in an automatic manner-A number of methods have been proposed
These include:
Global and local thresholding
Statistical approaches
Neural networks
Fuzzy logic
Thresholding of wavelet coefficients and related techniques.
Literature Review
Literature Review
Camilus et al.(2011)[1] propose an efficient method
To identify pectoral mussel using:
Watershed transformation
Merging algorithm to combine catchment basins
MIAS database(84 mammograms)
Literature Review
Literature Review
Pronoj et al.(2011)[2] reviews on :
Thresholding techniquesBoundary based methodHybrid techniquesWatershed transformationEdge detection:
SobelPrewittRobertsLaplacian of GaussianZero-crossCanny
Goal:oTo improve quality of imageoFacilate further processingoRemove noiseoRemove unwanted part from the background
Literature Review
Oliver et al.(2010)[3] worked on:
Local feature extraction from a bank of filters.
Performs training steps:-To automatically learn and select:
The features of microcalcifications.
Literature Review
Goal:oTo obtain different microcalcification morphology
Literature Review
Oliver et al.(2012)[4] :MC Detection based on:
microcalcifications morphologyLocal image features-
Set of feature is trained a pixel-based boosting classifier
Pixel-based boosting classifier:At each round automatically selects the most
salient microcalcifictions features.
Literature Review
Goal:oDetect microcalcification and cluster
Literature Review
Oliver et al. (2012)[4] :
Testing new mammogram:Only salient fractures are computed
Microcalcification clusters are found:By inspecting the local neighborhood of
each microcalcification.
Literature ReviewLiterature Review
Papadopoulus et al. (2008)[5] :
Microcalcification detection using neural network
Preprocessing image enhancement
Got best result by applying:The local range modification algorithmRedundant discrete wavelet linear stretching
and shrinkage algorithm.
Literature ReviewLiterature Review
Pal et al.(2008)[6] :To detect microcalcification cluster used:
oWeighted density function:-Position of microcalcifications
(take into account)Used:
oMulti-layered perception network for selecting 29 features
Features are used :- To segment mammograms
Literature ReviewLiterature Review
Razzi et al.(2009)[7] proposed :
A two-stage decomposition wavelet filteringFirst stage:Reduce background noise
Second stage:A hard thresholding technique:
-To identify microcalcificationCluster was considered if more then 3 microcalcifications were
detected in a 1cm2 area
Literature Review
Yu et al.(2010)[8] :
Clustered microcalcification detectionused combined :
-Model-based and statistical texture features
Firstly:Suspicious region containing microcalcification were detected using-
Wavelet filter and two thresholds
Literature Review
Yu et al. 2010 [8] proposed :
Secondly:Textural features were extracted:
-From each suspicious regionFeatures classified by:
-A back propagated neural network
Texture features based on both:oMorkov random fields and oFractal models
Literature Review
Wang et.al.(1989) [9]:
The mammograms are:-Decomposed into different frequency subbands.
The low-frequency subband discarded.
Literature Review
Literature Review
Daubechies I.(1992)[10]:
Wavelets are mainly used :
-Because of their dilation and translation properties-Suitable for non stationary signals.
Strickland et.at (1996)[11] :
Used biorthogonal filter bank-To compute four dyadic and -Two cinterpolation scales.
Applied binary threshold-operator -In six scales.
Literature Review
Heinlein et.al(2003)[12]:Goal: Enhancement of mammograms:
Derived The integrated wavelets:- From a model of microcalcifications
Literature Review
Zhibo et.al.(2007)[13]:A method aimed at minimizing image noise.
Optimize contrast of mammographic image featuresEmphasize mammographic features:
A nonlinear mapping function is applied:-To the set of coefficient from each level.
Use Contourlets:For more accurate detection of microcalcification clusters
The transformed image is denoised-using stein's thresholding [18].
The results presented correspond to the enhancement of regions with large masses only.
Literature Review
Fatemeh et.al.(2007) [14]:
Focus on:-Analysis of large masses instead of microcalcifications.- Detect /Classify mammograms:
Normal and Abnormal
Use Contourlets Transform:For automatic mass classification
Literature Review
Balakumaran et.al.(2010) [15] :
Focus on:- Microcalcification Detection
Use :- Wavelet Transform and Fuzzy Shell Clustering
Literature Review
Literature Review
Zhang et.al.(2013)[16] :
Use Hybrid Image Filtering Method:- Morphological image processing- Wavelet transform technique
Focus on:- Presence of microcalcification clusters
Literature Review
Lu et.al.(2013) [17]:
Use Hybrid Image Filtering Method:- Multiscale regularized reconstruction
Focus on:- Detecting subtle mass lesions in Digital breast
tomosynthesis (DBT)- Noise regularization in DBT reconstruction
Literature Review
Leeuw et.al.(2014) [18]:
Use:- Phase derivative to detect microcalcifications - A template matching algorithm was designed
Focus on:- Detect microcalcifications in breast
specimens using MRI - Noise regularization in image reconstruction
Literature Review
Shankla et.al.(2014)[19] :
Automatic insertion of simulated microcalcification clusters-in a software breast phantom
Focus on:-Algorithm developed as part of a virtual clinical trial (VCT) :
-Includes the simulation of breast anatomy, - Mechanical compression- Image acquisition- Image processing, displaying and interpretation.
Problem Statement
Burdensome Task Of Radiologist : Eye fatigue:
-Huge volume of images-Detection accuracy rate tends to decrease
Non-systematic search patterns of humansPerformance gap between :
Specialized breast imagers andgeneral radiologists
Interpretational Errors:Similar characteristics:
Abnormal and normal microcalcification
Problem Statement
Reason behind the problem( In real life):
The signs of breast cancer are:
Masses CalcificationsTumorLesionLump
Individual Research Areas
Problem Statement
Motivation to the Research
Motivation to the research: Goal
Better Cancer Survival Rates(Facilitate Early Detection ).
Provide “second opinion” : Computerized decisionsupport systems
Fast,Reliable, andCost-effective
QUICKLY AND ACCURATELY :Overcome the development of breast cancer
Challenges
Develop a logistic model:
Feature extraction Challenge:
-To determine the likelihood of CANCEROUS AREA -- From the image values of mammograms
Challenge:Occur in clusters
The clusters may vary in size from 0.05mm to 1mm in diameter.
Variation in signal intensity and contrast.May located in dense tissue
Difficult to detect.
Challenges
Database: mini-MIAS databasehttp://peipa.essex.ac.uk/pix/mias/
Class of Abnormality
Severity of Abnormality
The Location of The
Center of The
Abnormality and It’s
Diameter.
1 Calcification(25)
1.Benign(Calc-12)
2 Circumscribed Masses
3 Speculated Masses
4 Ill-defined Masses
5 Architectural Distortion
2.Malignant(Cancerous)
(Calc-13)
6 Asymmetry
7Normal
mdb223.jpg mdb226.jpg
mdb239.jpg mdb249.jpg
Figure01:X-ray image form mini-MIAS database
Database: Mini-MIAS Databasehttp://peipa.essex.ac.uk/pix/mias/
Mammography Image Analysis Society (MIAS) -An organization of UK research groups
• Consists of 322 images-- Contains left and right breast images for 161 patients
• Every image is 1024 X 1024 pixels in size
• Represents each pixel with an 8-bit word
• Reduced in resolution(Is not good enough for MC to be detectable)
•Very Poor Quality with .jpg compression effects(Original MIAS doesn’t have such artifacts)
Mini-MIAS Database
Mammography Image Analysis Society (MIAS) -An organization of UK research groups
Database: http://peipa.essex.ac.uk/pix/mias/
http://see.xidian.edu.cn/vipsl/database_Mammo.html
Plan of Action
Where Are We? Our Current Research Stage
Thesis SemesterM-3
Chart 01: Gantt Chart of this M.Sc thesis Showing the duration of task against the progression of time
Where Are We? Our Current Research Stage
Thesis SemesterM-3
Schematic representation of the system
Sche
mat
ic r
epre
sent
atio
n of
the
syst
em
Materials and Tools
Matlab 2014
Database: mini-MIAS
Removing Pectoral MuscleAnd
X-ray Label
Image SegmentationGoal: Removing X-ray Labeling And Pectoral muscles
Partitioning a digital image into multiple regions (sets of pixels).
GOAL OF SEGMENTATION:• To locate objects and boundaries (lines, curves, etc.) in
images.
• Result of image segmentation-A set of regions that collectively cover the entire image. (a) -A set of contours extracted from the image. (C)
• Each of the pixels in a region(1, 2, 3) are similar with respect to some characteristic or computed property, such as color, intensity, or texture.
• Adjacent regions(1, 2, 3) are significantly different with respect to the same characteristic(s).
Image Segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(intensity <130) (intensity >200)
1
2
3
(a) Segmentation Part
(C) Final Segmented Image
(b)Original image
Why Segmentation?
Image Segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
Proposed framework for breast profile segmentation
Plan of Action:
1. Original Image 2. Segmentation Part3. Final Segmented Image
4. Binary ImageLactiferous Sinus, Ducts, lobule
(After removing pectoral muscles, fatty tissues, Ligaments)
(intensity <130) (intensity >200)
Separating the Pectoral muscle
Image Segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
Keeping the biggest Cluster
(K-means clustering)(mdb256.jpg)
(a)Without Noise (b)With Noise
5.Final Segmented Image
BINARY Thresholding
For Two Different Ranges
1.Morphological Analysis:The Basic Operations are -
I.EROSIONII.DILATION
Using the basic operations we can perform -a)OPENINGb)CLOSING
Advanced Morphological Operation can then be implemented using Combinations Of All Of These
2.Image Smoothing/Filtering(Low pass):-Averaging (Drawback: Can vanish interesting details)
Lactiferous Sinus, Ducts, lobule(After removing pectoral muscles, fatty tissues, Ligaments)
5.Final Segmented Image
(a)Without Noise (b)With Noise
Techniques:
Noise Removing
More On Image Morphology Later
Image Morphology:-Deals with the shape (or morphology)of features in an image-Operate on bi-level images
Structuring Elements, Hits & Fits
B
AC
Structuring Element
Fit: All on pixels in the structuring element cover on pixels in the image
Hit: Any on pixel in the structuring element covers an on pixel in the image
All morphological processing operations are based on these simple ideas
Image Morphology Noise Removing
Structuring elements can be any size and make any shape
However, for simplicity we will use rectangular structuring elements with their origin at the middle pixel
1 1 1
1 1 1
1 1 1
0 0 1 0 00 1 1 1 01 1 1 1 10 1 1 1 00 0 1 0 0
0 1 0
1 1 1
0 1 0
Structuring Elements, Hits & Fits
Image Morphology Noise Removing
0 0 0 0 0 0 0 0 0 0 0 00 0 0 1 1 0 0 0 0 0 0 00 0 1 1 1 1 1 0 0 0 0 00 1 1 1 1 1 1 1 0 0 0 00 1 1 1 1 1 1 1 0 0 0 00 0 1 1 1 1 1 1 0 0 0 00 0 1 1 1 1 1 1 1 0 0 00 0 1 1 1 1 1 1 1 1 1 00 0 0 0 0 1 1 1 1 1 1 00 0 0 0 0 0 0 0 0 0 0 0
B C
A
1 1 1
1 1 1
1 1 1Structuring Element 1
0 1 0
1 1 1
0 1 0Structuring Element 2
Structuring Elements, Hits & Fits
Image Morphology Noise Removing
•The structuring element is moved across every pixel in the original image to give a pixel in a new processed image(very like spatial filtering)
•The value of this new pixel depends on the operation performed
•There are two basic morphological operations:
Erosion and Dilation
Structuring Elements, Hits & Fits
Image Morphology Noise Removing
Erosion of image f by structuring element s is given by f sThe structuring element s is positioned with its origin at (x, y) and the new pixel value is determined using the rule:
Erosion
=otherwise 0
fits if 1),(
fsyxg
Structuring Elements, Hits & Fits
A morphological opening of an image is an erosion followed by a dilation
Noise Removing 1. Morphological Analysis
What Is Erosion For?
Erosion can split apart joined objects
Erosion can split apart
%noise removingse = strel('disk',25);for i=1:19erode_bolb = imerode(largest_bolb,se);end
Original imageErosion by
3*3square
structuring element
Erosion by 5*5
square structuring
element
Noise Removing 1. Morphological Analysis
Watch out: Erosion shrinks objects
Erosion Example
Structuring Element
Original Image Processed Image With Eroded Pixels
Noise Removing 1. Morphological Analysis
Erosion Example
Structuring Element
Original Image Processed Image
Dilation
Image Morphology X-ray Label Removing
Dilation of image f by structuring element s is given by f sThe structuring element s is positioned with its origin at (x, y) and the new pixel value is determined using the rule:
⊕
=otherwise 0
hits if 1),(
fsyxg
A morphological closing of an image is a dilation followed by an erosion
bw_image = im2bw(Binary_image);imtool(bw_image)se1 = strel ('line', 3,0);se2 = strel ('line', 3,90);
for i=1:9BW2= imdilate (bw_image, [se1
se2], 'full')BW2 = imfill(BW2,'holes');
end
Noise Removing 1. Morphological Analysis
Structuring Elements, Hits & Fits
Structuring Element
Original Image Processed Image
Dilation Example
Noise Removing 1. Morphological Analysis
Dilation Example
Structuring Element
Original Image Processed Image With Dilated Pixels
Dilation ExampleOriginal image
Hole fillingInside the blob(dilation)
Result image(Label Removed)
mdb240.jpg
Binary image
A morphological closing of an image is an dilation followed by a erosion
%hole filling with in the bolbse = strel('disk',39);for i=1:19closeBW_largest_bolb = imclose(largest_bolb,se);
After Removing Some NoiseImage Containing Noise(mdb041.jpg)
Noise Removing 2.Image Smoothing/Filtering(Low pass):
After Removing Some NoiseImage Containing Noise(mdb041.jpg)
Noise Removing
Chosen Technique 2D MEDIAN FILTERING FOR SALT AND PEPPER NOISE
I = medfilt2(I, [1 5]); Median filtering is a nonlinear operation often used in image processing to reduce "salt and pepper" noise. A median filter is more effective than convolution when the goal is to simultaneously reduce noise and preserve edges. Since all the mammograms are in high quality images, there is no need to perform median filtering
Why choosing?
2.Image Smoothing/Filtering(Low pass):
1. Morphological AnalysisOVER
-Does not work will on all the image [I = medfilt2(I, [1 5]);] •No effect most of the time•Absence of salt and peeper noise
-Tendency of loosing interesting details
Class: Benign
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
mdb212 150 200mdb214 150 200mdb218 150 210mdb219 150 210mdb222 150 210mdb223 150 210mdb226 150 210mdb227 150 210mdb236 150 210mdb240 150 210mdb248 150 210mdb252 140 210
(inte
nsity
<15
0) (
inte
nsity
>20
0)
1
2
3
(b) Segmentation Part
(C) Final Segmented Image
(a)Original image
mdb236.jpg
(b)Segmentation Part (c) Final Segmented Image(a)Main Image (e)Image containing Duct, Lobules, Sinus
mdb001.jpg
mdb254.jpg
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(d)Binary Image
Achievement: X-Ray Label removedClass: Benign
(b)Segmentation Part (c) Final Segmented Image(a)Main Image (e)Image containing only Pectoral muscle
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(d)Binary Image
What we need
mdb212.jpg
Issues 1.Biggest Cluster Does Not Contain BreastProduce Artifacts In Pectoral muscle And Breast Region
What we have
Class: Benign
(b)Segmentation Part (c) Final Segmented Image(a)Main Image(e)Image Containing Only Pectoral muscle
mdb001.jpg
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(d)Binary Image
mdb214.jpg
mdb218.jpg
What We Need What We Have
Issues 1.Biggest Cluster Does Not Contain BreastProduce Artifacts In Pectoral muscle And Breast Region
Class: Benign
(b)Segmentation Part (c) Final Segmented Image(a)Main Image(e)Image containing duct, lobules,
sinus & Pectoral muscle
mdb001.jpg
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(d)Binary Image
mdb222.jpg
mdb223jpg
mdb226jpg
Wha
t we
wan
t What w
e Have
Issues 1.Biggest Cluster Does Not Contain BreastProduce Artifacts In Pectoral muscle And Breast Region
Class: Benign
(b)Segmentation Part (c) Final Segmented Image(a)Main Image
mdb001.jpg
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(d)Binary Image
mdb240.jpg
mdb248.jpg
mdb252.jpg
Wha
t we
wan
t What w
e Have
(e)Image containing duct, lobules, sinus & Pectoral musc
Issues 1.Biggest Cluster Does Not Contain BreastProduce Artifacts In Pectoral muscle And Breast Region
Class: Benign
Class: Malignant
mdb209 140 210mdb211 140 210mdb213 140 210mdb216 140 210mdb231 140 210mdb233 140 210mdb238 140 210mdb239 140 210mdb241 140 210mdb245 140 210mdb249 140 210mdb253 140 210mdb254 140 210mdb256 140 210
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(inte
nsity
<14
0) (
inte
nsity
>21
0)
1
23
(b)Segmentation Part
(C) Final Segmented Image
(a)Original imagemdb209.jpg
(a) Original image
mdb241.jpg
(b)Segmentation Part (c) Final Segmented Image(a)Main Image(e)Image containing duct, lobules, sinus
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
(d)Binary Image
Class: Malignant
Achievement: X-Ray Label removed
mdb238.jpg
mdb245.jpg
Labe
l Rem
oved
(a)Main Image
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
mdb212.jpg
Issues Level Remain In The ImageProduce Artifacts In Pectoral muscle And Breast Region
Class: Malignant
mdb209.jpg
(b)Segmentation Part
(c) Final Segmented Image
(d)Binary Image
(e)Image Containingonly label
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
mdb212.jpg
Issues Pectoral muscle Remain In The Image
Produce Artifacts In Breast Region
Class: Malignant
(b)Segmentation Part (c) Final Segmented Image(a)Main Image(e)Image containing only pectoral muscle(d)Binary Image
mdb216.jpg
mdb213.jpg
(b)Segmentation Part (c) Final Segmented Image(a)Main Image(e)Image containing only pectoral muscle(d)Binary Image
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
Expected output: PECTORAL muscle, DUCT, LOBULES, SINUS,LIGAMENTS
mdb256.jpg
Output ImageMain image
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
Issues
Missing part Main image Image containing Output Image
Fatty tissue area,ligaments
Duct, Lobules, Sinus
Fatty tissue area,Duct, Lobules, Sinus,
ligaments
X-ray Labels
Fatty tissue area,Duct, Lobules, Sinus,
ligaments
Pectoral muscle
mdb001.jpg
mdb209.jpg
mdb213.jpg
Challenge
Find the binary threshold values.
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
2. Segmentation Part
(intensity <150) (intensity >200)
(K-means clustering)
1.Need A Non-supervised Method
mdb212 150 200
mdb214 150 200
mdb218 150 210
mdb219 150 210
mdb222 150 210
mdb223 150 210
mdb226 150 210
mdb227 150 210
mdb236 150 210
mdb240 150 210
mdb248 150 210
mdb252 140 210
No pre-defined threshold value
Image segmentation K-means Clustering
Goal: Removing X-ray Labeling And Pectoral muscles
Figure: Internal breast structure
2.Keeping fatty tissues and ligamentsChallenge
mdb001.jpg
mdb212.jpg
mdb223jpg
mdb238.jpg
mdb209.jpg
(a)Main Image (b)Result Image
(a)Main Image (b)Result Image
X-ray Label Removing Finding The Big BLOB
The types of noise :High Intensity Rectangular LabelLow Intensity LabelTape Artifacts
1.Binarizatin of original image.
2.Find the biggest blob.
Plan of Action:
function [outim] = bwlargestblob( im,connectivity)if size(im,3)>1,
error('bwlargestblob accepts only 2 dimensional images');end
[imlabel totalLabels] = bwlabel(im,connectivity);sizeBlob = zeros(1,totalLabels);for i=1:totalLabels,
sizeblob(i) = length(find(imlabel==i));end[maxno largestBlobNo] = max(sizeblob);
outim = zeros(size(im),'uint8');outim(find(imlabel==largestBlobNo)) = 1;
end
img=im2bw(img);(threshold luminance level-=0.5)
X-ray Label Removing Finding The Big BLOB
1.Binarizatin of original image.
2.Find the biggest blob.
Plan of Action:
(threshold luminance level-=0.5)
Original image Binary Image
mdb219.jpg
(a) Artifacts (Hole) in ROI
(b)Absence of Ligaments and fatty tissue
mdb231.jpgmdb253.jpg
(c) Absence of pectoral muscles
Original image Binary Image
Label successfully removed Issues
X-ray Label Removing Finding The Big BLOB
Original image Binary Image(threshold luminance level-=0.5)
mdb212.jpg
mdb214.jpg
mdb218.jpg
Original image Binary Image(threshold luminance level-=0.5)
mdb219.jpg
mdb222.jpg
mdb223.jpg
Class: Benign Issue with fatty tissues and ligaments existence
X-ray Label Removing Finding The Big BLOB
Original image Binary Image(threshold luminance level-=0.5) Original image Binary Image
(threshold luminance level-=0.5)
mdb226.jpg
mdb227.jpg
mdb236.jpg
mdb240.jpg
mdb248.jpg
mdb252.jpg
Class: Benign Issue with fatty tissues and ligaments existence
X-ray Label Removing Finding The Big BLOB
Original imageBinary Image
(threshold luminance level-=0.5) Original image Binary Image(threshold luminance level-=0.5)
Issue with fatty tissues and ligaments existenceClass: Malignant
mdb209.jpg
mdb211.jpg
mdb213.jpg
mdb216.jpg
mdb231.jpg
mdb233.jpg
X-ray Label Removing Finding The Big BLOB
Original image Binary Image(threshold luminance level-=0.5) Original image Binary Image
(threshold luminance level-=0.5)
Issue with fatty tissues and ligaments existenceClass: Malignant
mdb238.jpg
mdb239.jpg
mdb241.jpg
mdb245.jpg
mdb249.jpg
mdb253.jpg
mdb256.jpg
X-ray Label Removing Finding The Big BLOB
Moving towards solution
Issue With Fatty Tissues And Ligaments Existence
X-ray Label Removing
Plan of Action:
1.Binarize the image
2.Fill inside the hole region of the binary image
3.Finding the largest Blob:
4.Keep the Largest Blob and discard other blobs(to remove X-ray level)
function [outim] = bwlargestblob( im,connectivity)if size(im,3)>1,
error('bwlargestblob accepts only 2 dimensional images');end
[imlabel totalLabels] = bwlabel(im,connectivity);sizeBlob = zeros(1,totalLabels);for i=1:totalLabels,
sizeblob(i) = length(find(imlabel==i));end[maxno largestBlobNo] = max(sizeblob);
outim = zeros(size(im),'uint8');outim(find(imlabel==largestBlobNo)) = 1;
X-ray Label Removing
Image Morphology
Experimental results:
Goal: Region filling(Region inside the blob)
Original imageFinding biggest blob
(Level removed)Hole filling
Inside the blob(dialation)Result image
(Label Removed)
mdb240.jpg
mdb219.jpg
mdb231.jpg
Binary image
Direct Binarization Without Image enhancement
X-ray Label Removing
Experimental results:
Original imageResult image
(Label Removed)
mdb240.jpg
Issues
mdb219.jpg
mdb231.jpg
Binary image
1.Does not always produce appealing output
2.Some details are missing(Details around Edge region )
Image MorphologyGoal: Region filling(Region inside the blob)
X-ray Label Removing
Direct Binarization Without Image enhancement
Original image Result image (Label Removed)
mdb240.jpg
mdb219.jpg
mdb231.jpg
Issues
1.Does not always produce appealing output
2.Some details are missing(Details around Edge region )
mdb212.jpg
mdb214.jpg
mdb219.jpg
mdb226.jpg
Experimental results:
Image MorphologyGoal: Region filling(Region inside the blob)
X-ray Label Removing
Direct Binarization Without Image enhancement
-To find largest blobUse -Otsu’s thresholding technique (graytrash) [20]
-Finding Bi-level the image(im2bw)
To Achieve The Desired Final Result:
-ApplyA Range Of Techniques on original image
[20] Otsu, N., "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.
X-ray Label Removing
X-ray Label Removing
1. Histogram equalization of the original X-ray image2. Adjust image contrast3. Apply Otsu's Thresholding Method [20] and
find bi-level the image which has several blobs in it. 4. Finding the Largest blob (Bwlargest.bolb)5. Hole filling within the blob region6. Keep the true pixel value covering only the area of largest
blob and discard other features from the original image7. X-ray label is successfully removed
Plan of Action
[20] Otsu, N., "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.
1.Original image
2.HistogramEqualization
3.Contrast Image
4.Binary Image
mdb239.jpg
Combining Range of techniques
J = histeq(I); %histogram equalization
contrast_image = imadjust(J, stretchlim(J), [0 1]); %high contrast image
%Apply Thresholding to the Image level = graythresh(contrast_image);
%GRAYTHRESH Global image threshold using %Otsu's methodbw_image = im2bw(contrast_image, level);%getting binary image
X-ray Label Removing
5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
Combining Range of techniquesX-ray Label Removing
Result image(Label Removed)
Original image
Compare the original and final image
X-ray Label Removing
Experimental results
X-ray Label Removing
X-ray Label Removing1.Original image
2.HistogramEqualization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
mdb212.jpg
mdb214.jpg
mdb214.jpg
mdb218.jpg
mdb219.jpg
Benign
X-ray Label Removing Benign1.Original image
2.HistogramEqualization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
mdb222.jpg
mdb223.jpg
mdb226jpg
mdb227jpg
X-ray Label Removing Benign1.Original image
2.HistogramEqualization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
mdb226.jpg
mdb240.jpg
mdb248.jpg
mdb252.jpg
X-ray Label Removing Malignant1.Original image
2.HistogramEqualization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
mdb209.jpg
mdb211.jpg
mdb213.jpg
mdb216.jpg
mdb231.jpg
X-ray Label Removing Malignant1.Original image
2.HistogramEqualization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
mdb233.jpg
mdb238.jpg
mdb239.jpg
mdb241.jpg
X-ray Label Removing Malignant1.Original image
2.HistogramEqualization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole fillingInside the blob
7.Result image(Label Removed)
mdb245.jpg
mdb249.jpg
mdb253.jpg
mdb256.jpg
Successful
X-ray Label Removing
Finally!
Removing pectoral muscleKeeping fatty tissues and ligaments
mdb212.jpg(a)Main Image (b)Result Image
mdb213.jpg(a)Main Image (b)Pectoral Muscle
mdb214.jpg
Main Image
Result Image
o Fat t y t i s s ue are ao Duc to Lobul e so Si nuso l i gam e nt s
Extraction of ROIRemoving pectoral muscle
Why removing pectoral muscle?
o Pectoral muscle will never contain micro-calcification
o Less Computational Time And Cost-Operation on small image area
Existence of micro-calcification:
ROI
Edge Detection of pectoral muscleRemoving pectoral muscle
Possible Approach To Edge-detection:
1.Scanning pixel value intensity at each points2.find out the sudden big intensity change at the edge location3.Mark the pixels at edge location4.Estimate a straight line depending on the marked edge points
Approach-01:
Problem faced in Approach-01:-Finding appropriate Thresholding value is an unsupervised method,
which will work on every image -The threshold value must be found in an unsupervised manner-Any predefined threshold value will not produce desired output for all image
mdb212 150 200mdb214 130 205mdb218 150 210mdb219 120 200mdb222 150 210mdb223 150 225mdb226 110 210mdb227 150 230mdb236 160 210mdb240 150 200mdb248 150 210mdb252 140 210
Edge Detection of pectoral muscleRemoving pectoral muscle
Possible Approach To Edge-detection:
1.Segment the image
2.Separate the pectoral muscle form the Duct, Lobules, Sinus region Making all the pixels black(zero)resides in the fatty tissue and ligament area
3.Find the binary image of image found in step 2(it will be used as outer image)
4.Erode the image found in step-3 (it will be used as inner image)
5.Subsract the inner image from the outer image to get the edge
Approach-02:
Visualization in next slide
Edge Detection of pectoral muscleRemoving pectoral muscle
1.Original imagemdb212.jpg
2.Segmentation Part
3.Fatty tissue& Ligament removed
Possible Approach To Edge-detection(Approach-02):
Edge Detection of pectoral muscleRemoving pectoral muscle
4.Binary Version(outer)
5.Binary Version(inner)
6.Edge(outer-inner)
Possible Approach To Edge-detection(Approach-02):
Experimental results
Edge Detection of pectoral muscleRemoving pectoral muscle
Possible Approach To Edge-detection(Approach-02):
Edge Detection of pectoral muscleRemoving pectoral muscle
mdb212.jpg
mdb214.jpg
mdb218.jpg
mdb252.jpg
1.Original image 2.Segmentation Part 4.Binary Version(outer)
3.Fatty tissue& Ligament removed 5.Binary Version(inner) 6.Edge(outer-inner)
Edge Detection of pectoral muscleRemoving pectoral muscle
mdb223.jpg
mdb226.jpg
mdb240.jpg
mdb248.jpg
1.Original image 2.Segmentation Part3.Fatty tissue
& Ligament removed4.Binary Version(outer)5.Binary Version(inner) 6.Edge(outer-inner)
Edge Detection of pectoral muscleRemoving pectoral muscle
1.Pectoral muscle and ligaments in fatty tissue area got merged
mdb218.jpg
mdb240.jpg
1.Original image 2.Segmentation Part3.Fatty tissue
& Ligament removed4.Binary Version(outer)5.Binary Version(inner) 6.Edge(outer-inner)
Problems faced in (Approach-02):
Edge Detection of pectoral muscleRemoving pectoral muscle
2.Discontinuity in Pectoral muscle edge
mdb252.jpg
mdb226.jpg
mdb252.jpg
mdb248.jpg
1.Original image 2.Segmentation Part3.Fatty tissue
& Ligament removed4.Binary Version(outer)5.Binary Version(inner) 6.Edge(outer-inner)
Problems faced in (Approach-02):
Edge Detection of pectoral muscleRemoving pectoral muscle
Problems faced in (Approach-02):
3.Same thresholding value(i.e.,130-210,) does not work well on all the images and Produce improper output(complete black image as output)
1.Original image 2.Segmentation Part3.Fatty tissue
& Ligament removed4.Binary Version(outer)
5.Binary Version(inner)6.Edge(outer-inner)
Edge Detection of pectoral muscleRemoving pectoral muscle
Points to be noted from approach-2:
-Pectoral muscle a Triangular areamdb212.jpg
mdb214.jpg
Based on this point: Moving on to approach -03
mdb209.jpg
(2)Binary Image(1)Original Image
Triangle Detection of pectoral muscleRemoving pectoral muscle
1.Fing the triangular area of the pectoral muscle region
I. Finding white seeding pointII. Finding the 1st black point of 1st row after getting a white seeding pointIII. Draw a horizontal line in these two points.IV. finding the 1st black point of 1st column after getting a white seeding pointV. Draw a vertical line and angular line.
2.Making all the pixels black(zero)resides in the pectoral muscle area
Approach-03(Triangle Detection of pectoral muscle):
Visualization in next slide
Triangle Detection of pectoral muscleRemoving pectoral muscle
Approach-03(Triangle Detection of pectoral muscle):
mdb212.jpg1.Original image
2.Contrast stretching
3.Binary of contrast image
stratching_in_range=uint8(imadjust(I,[0.01 0.7],[1 0]));
BW=~stratching_in_range;
Triangle Detection of pectoral muscleRemoving pectoral muscle
Approach-03(Triangle Detection of pectoral muscle):
4.Triangle
5.Triangle Filled
6.muscle removed
Experimental results
Removing pectoral muscleApproach-03(Triangle Detection of pectoral muscle):
Triangle Detection of pectoral muscle
Triangle Detection of pectoral muscleRemoving pectoral muscle
mdb212.jpg
mdb214.jpg
1.Original image 2.Contrast stretching 3.Binary of contrast image 4.Triangle
mdb240.jpg
mdb248.jpg
5.Triangle Filled 6.muscle removed
Class: Benign
Triangle Detection of pectoral muscleRemoving pectoral muscle
mdb222.jpg
mdb226.jpg
mdb227.jpg
2.Contrast stretching1.Original image 3.Binary of contrast image 4.Triangle
Problems faced in (Approach-03):
5.Triangle Filled 6.muscle removed
The triangle does not always indicates the proper pectoral muscle area.Reason: Discontinuity in edges (First 3 or 4 rows and columns)
it is caused by artifacts in mammogram
Class: Benign
Triangle Detection of pectoral muscleRemoving pectoral muscle
Problems faced in (Approach-03):
mdb218.jpg
1.Original image 2.Contrast stretching 3.Binary of contrast image 4.Triangle
mdb219.jpg
5.Triangle Filled 6.muscle removed
mdb218.jpg
The triangle does not always indicates the proper pectoral muscle area.Reason: Discontinuity in edges (First 3 or 4 rows and columns)
Class: Benign
Triangle Detection of pectoral muscleRemoving pectoral muscle
Problems faced in (Approach-03):
1.Original image 2.Contrast stretching 3.Binary of contrast image 4.Triangle
mdb222.jpg
mdb219.jpg
5.Triangle Filled 6.muscle removed
The triangle does not always indicates the proper pectoral muscle area.Reason: Discontinuity in edges (First 3 or 4 rows and columns)
Class: Benign
mdb223.jpg
2.Contrast stretching1.Original image 3.Binary of contrast image 4.Triangle 5.Triangle Filled 6.muscle removed
Triangle Detection of pectoral muscleRemoving pectoral muscle
Problems faced in (Approach-03):
Defects in mammogram
Class: Benign
2.Contrast stretching1.Original image
3.Binary of contrast image 4.Triangle
5.Triangle Filled 6.muscle removed
Triangle Detection of pectoral muscleRemoving pectoral muscle
Problems faced in (Approach-03):
Defects in mammogram
mdb227.jpg
Class: Benign
Triangle Detection of pectoral muscleRemoving pectoral muscle
1.Original image 2.Contrast stretching 3.Binary of contrast image 4.Triangle 5.Triangle Filled 6.muscle removed
Class: Malignant
mdb241jpg
Mdb249.jpg
mdb211.jpg
Problems faced in (Approach-03): 2.Discontinuity in edge lines causes false output
Triangle Detection of pectoral muscleRemoving pectoral muscle
1.Original image
2.Contrast stretching
3.Binary of contrast image
4.Triangle
5.Triangle Filled
6.muscle removed
Class: Malignant
mdb256.pg
Triangle Detection of pectoral muscleRemoving pectoral muscle
mdb212.jpg
mdb214.jpg
1.Original image 2.Contrast stretching 3.Binary of contrast image 4.Triangle
mdb240.jpg
mdb248.jpg
5.Triangle Filled 6.muscle removed
Class: Benign
Successful
Pectoral Muscle Removing
Finally!
Improved Computer Assisted Screening
Enhancement of digitized mammogram
Goal
MAIN NOVELTY
Input image
BandpassDirectionalsubbands
BandpassDirectionalsubbands
Main Novelty
-Contourlet Transform
- Specific Edge Filter (Prewitt Filter):To enhance the directional structures of the image in
the contourlet domain.
- Recover an approximation of the mammogram (with the microcalcifications enhanced):
Inverse contourlet transform is applied
Details in upcoming slides
Based on the classical approach used in transform methods for image processing.
1. Input mammogram
2. Forward CT
3. Subband Processing
4. Inverse CT
5. Enhanced Mammogram
Schematic representation of the system
Contourlet transformation
Implementation Based On :
• A Laplacian Pyramid decomposition followed by -
• Directional filter banks applied on each band pass sub-band.
The Result Extracts:-Geometric information of images.
Details in upcoming slides
Main Novelty
Enhancement of the Directional Subbands
The Contourlet TransformLaplacian Pyramid: 3 level
DecompositionFrequency partitioning of a directional filter bank
Decomposition level l=3
The real wedge-shape frequency band is 23=8.
horizontal directions are corresponded by sub-bands 0-3
Vertical directions are represented by sub-bands 4-7
Details in upcoming slides
Enhancement of the Directional Subbands
The Contourlet TransformLaplacian Pyramid: 3 level
Decomposition
Laplacian Pyramid Level-1
Laplacian Pyramid Level-2
Laplacian Pyramid Level-3
8 Direction
4 Direction
4 Direction
(mdb252.jpg)
Enhancement of the Directional Subbands
The Contourlet TransformLaplacian Pyramid: 3 level
Decomposition
Wedge-shape frequency band is 23=8.
Horizontal directions are corresponded by sub-bands 0-3
(1) sub-band 0
(2) sub-band 1
(3) sub-band 2(4) sub-band 3
Contourlet coefficient at level 4
Enhancement of the Directional Subbands
The Contourlet TransformLaplacian Pyramid: 3 level
DecompositionContourlet coefficient at level 4
Wedge-shape frequency band is 23=8.
Vertical directions are represented by sub-bands 4-7
(5) sub-band 4
(6) sub-band 5
(7) sub-band 6
(8) sub-band 7
Enhancement of the Directional Subbands
The Contourlet TransformLaplacian Pyramid: 3 level
Decomposition
(a) Main Image(mdb252.jpg)
(b) Enhanced Image(Average in all 8 direction)
(a) Main image(Toy Image)
Contourlet Transform Example
(b) Horizontal Direction
(c) Vertical Direction
Directional filter banks: Horizontal and Vertical
Contourlet Transform ExampleDirectional filter banks
Horizontal directions are corresponded by sub-bands 0-3
(1) sub-band 0
(2) sub-band 1
(3) sub-band 2
(4) sub-band 3
Contourlet Transform ExampleDirectional filter banks
Vertical directions are represented by sub-bands 4-7
(5) sub-band 4
(6) sub-band 5
(7) sub-band 6
(8) sub-band 7
Why Contourlet?
Why Contourlet?
•Decompose the mammographic image:-Into directional components:
To easily capture the geometry of the image features.
Details in upcoming slides
Target
Details in upcoming slides
• This decomposition offers:
-Multiscale localization(Laplacian Pyramid) and -A high degree of directionality and anisotropy.
Why Contourlet? Usefulness of Contourlet
Directionality:Having basis elements Defined in variety of directions
Anisotrophy:Basis Elements having Different aspect ration
Contourlet Transform Concept
(a)Wavelet(Require a lot of dot for fine resolution)
(b)Contourlet(Requires few different elongated shapes
in a variety of direction following the counter)
3 Different Size of Square Shape brush stroke(Smallest, Medium, Largest) to provide Multiresolution Image
Example: Painter Scenario
Why Contourlet?
2-D Contourlet Transform (2D-CT) Discrete WT
Handles singularities such as edges in a more powerful way
Has basis functions at many orientations has basis functions at three orientations
Basis functions appear a several aspectratios
the aspect ratio of DWT is 1
CT similar as DWT can beimplemented using iterative filter banks.
Advantage of using 2D-CT over DWT:
Details in upcoming slides
Input image
BandpassDirectionalsubbands
BandpassDirectionalsubbands
Plan-of-Action
For microcalcifications enhancement :
We use-The Contourlet Transform(CT) [12]
The Prewitt Filter.
12. Da Cunha A. L., Zhou J. and Do M. N,: The Nonsubsampled Contourlet Transform: Theory, Design, and Applications, IEEE Transactions on Image Processing,vol. 15, (2006) pp. 3089-3101
Art-of-Action
An edge Prewitt filter to enhance the directional structures
in the image.
Contourlet transform allows decomposing the image in
multidirectional and multiscale subbands[21].
21. Laine A.F., Schuler S., Fan J., Huda W.: Mammographic feature enhancement by multiscale analysis, IEEE Transactions on Medical Imaging, 1994, vol. 13, no. 4,(1994) pp. 7250-7260
This allows finding • A better set of edges,• Recovering an enhanced mammogram with better visual characteristics.
Microcalcifications have a very small size a denoising stage is not implemented
in order to preserve the integrity of the injuries.
Decompose the digital mammogram
Using Contourlet transform
(b) Enhanced image(mdb238.jpg)
(a) Original image (mdb238.jpg)
Method
CT is implemented in two stages:
1. Subband decomposition stage
2. Directional decomposition stages.
Details in upcoming slides
Method
1. Subband decomposition stage
For the subband decomposition:- The Laplacian pyramid is used [22]
Decomposition at each step:-Generates a sampled low pass version of the original-The difference between :
The original image and the prediction.
22. Park S.-I., Smith M. J. T., and Mersereau R. M.: A new directional Filter bank for image analysis andclassification, Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP '99), vol.3, (1999) pp. 1417-1420
Details ……..
Method
1. Subband decomposition stage
Details ……..
1. The input image is first low pass filtered
2. Filtered image is then decimated to get a coarse(rough) approximation.
3. The resulting image is interpolated and passed through Synthesis filter.
4. The obtained image is subtracted from the original image :To get a bandpass image.
5. The process is then iterated on the coarser version (high resolution)of the image.
Plan of Action
Method
2.Directional Filter Bank (DFB)
Details ……..
Implemented by using an L-level binary tree decomposition :resulting in 2L subbands
The desired frequency partitioning is obtained by :Following a tree expanding rule
- For finer directional subbands [22].
22. Park S.-I., Smith M. J. T., and Mersereau R. M.: A new directional Filter bank for image analysis andclassification, Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP '99), vol.3, (1999) pp. 1417-1420
The Contourlet Transform
The CT is implemented by:Laplacian pyramid followed by directional filter banks (Fig-01)
Input image
BandpassDirectionalsubbands
BandpassDirectionalsubbands
Figure 01: Structure of the Laplacian pyramid together with the directional filter bank
The concept of wavelet:University of Heidelburg
The CASCADE STRUCTURE allows:- The multiscale and
directional decomposition to be independent
- Makes possible to:Decompose each scale into
any arbitrary power of two's number of directions(4,8,16…)
Figure 01
Details ………….
Decomposes The Image Into Several Directional Subbands And Multiple Scales
Figure 02: (a)Structure of the Laplacian pyramid together with the directional filter bank(b) frequency partitioning by the contourlet transform(c) Decomposition levels and directions.
(a) (b)
Input image
BandpassDirectionalsubbands
BandpassDirectionalsubbands
Details….
(c)
DenoteEach subband by yi,j
Wherei =decomposition level and J=direction
The Contourlet TransformDecomposes The Image Into Several Directional Subbands And Multiple Scales
The processing of an image consists on:-Applying a function to enhance the regions of
interest.
In multiscale analysis:
Calculating function f for each subband :-To emphasize the features of interest-In order to get a new set y' of enhanced subbands:
Each of the resulting enhanced subbands can be expressed using equation 1.
)(', , jiyfjiy = ………………..(1)
-After the enhanced subbands are obtained, the inverse transform is performed to obtain an enhanced image.
Enhancement of the Directional Subbands
The Contourlet Transform
Denote
Each subband by yi,jWherei =decomposition level and J=direction Details….
Enhancement of the Directional Subbands
The Contourlet Transform
Details….
The directional subbands are enhanced using equation 2.
=)( , jiyf)2,1(
,1 nnWjiy
)2,1(,2 nnWjiy
If bi,j(n1,n2)=0
If bi,j(n1,n2)=1………..(2)
Denote
Each subband by yi,jWherei =decomposition level and J=direction
W1= weight factors for detecting the surrounding tissueW2= weight factors for detecting microcalcifications
(n1,n2) are the spatial coordinates.
bi;j = a binary image containing the edges of the subband
Weight and threshold selection techniques are presented on upcoming slides
Enhancement of the Directional Subbands
The Contourlet Transform
The directional subbands are enhanced using equation 2.
=)( , jiyf)2,1(
,1 nnWjiy
)2,1(,2 nnWjiy
If bi,j(n1,n2)=0
If bi,j(n1,n2)=1………..(2)
Binary edge image bi,j is obtained :-by applying an operator (prewitt edge detector)
-to detect edges on each directional subband.
In order to obtain a binary image:A threshold Ti,j for each subband is calculated.
Details….
Weight and threshold selection techniques are presented on upcoming slides
Threshold Selection
The Contourlet Transform
Details….
The microcalcifications appear :
On each subband Over a very
homogeneous background.
Most of the transform coefficients:
-The coefficients corresponding to theinjuries are far from background value.
A conservative threshold of 3σi;j is selected:where σi;j is the standard deviation of the corresponding subband y I,j .
Weight Selection
The Contourlet Transform
Exhaustive tests:-Consist on evaluating subjectively a set of 322 different mammograms
-With Different combinations of values,
The weights W1, and W2 are determined:-Selected as W1 = 3 σi;j and W2 = 4 σi;j
These weights are chosen to:keep the relationship W1 < W2:
-Because the W factor is a gain -More gain at the edges are wanted.
Experimental Results
Applying Contourlet Transformation Benign
Original image Enhanced image
Goal: Microcalcification Enhancement
mdb222.jpg
mdb223.jpg
Original image Enhanced image
mdb248.jpg
mdb252.jpg
Applying Contourlet Transformation Benign
Original image Enhanced image
mdb226.jpg
mdb227.jpg
Original image Enhanced image
mdb236.jpg
mdb240.jpg
Goal: Microcalcification Enhancement
Applying Contourlet Transformation Benign
Original image Enhanced image Original image Enhanced image
mdb218.jpgmdb219.jpg
Goal: Microcalcification Enhancement
Applying Contourlet Transformation MalignantGoal: Microcalcification Enhancement
Original image Enhanced image
mdb209.jpg
mdb211.jpg
Original image Enhanced image
mdb213.jpg
mdb231.jpg
Applying Contourlet Transformation MalignantGoal: Microcalcification Enhancement
Original image Enhanced image
mdb238.jpg
mdb239.jpg
Original image Enhanced image
mdb241.jpg
mdb249.jpg
Original image Enhanced image
mdb253.jpg
Original image Enhanced image
Applying Contourlet Transformation MalignantGoal: Microcalcification Enhancement
mdb256.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb003.jpg
mdb004.jpg
Original image Enhanced image
mdb006.jpg
mdb007.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb009.jpg
mdb018.jpg
Original image Enhanced image
mdb027.jpg
mdb033.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb046.jpg
mdb056.jpg
Original image Enhanced image
mdb060.jpg
mdb066.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb070.jpg
mdb073.jpg
Original image Enhanced image
mdb074.jpg
mdb076.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb093.jpg
mdb096.jpg
Original image Enhanced image
mdb101.jpg
mdb012.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb128.jpg
mdb137.jpg
Original image Enhanced image
mdb146.jpg
mdb154.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb166.jpg
mdb169.jpg
Original image Enhanced image
mdb224.jpg
mdb225.jpg
Applying Contourlet Transformation NormalGoal: Microcalcification Enhancement
Original image Enhanced image
mdb263.jpg
mdb294.jpg
Original image Enhanced image
mdb316.jpg
mdb320.jpg
Wavelet Transformation
Use Separable Transform
2D Wavelet Transform
Visualization
Label ofapproximation
HorizontalDetails
HorizontalDetails
VerticalDetails
DiagonalDetails
VerticalDetails
DiagonalDetails
Use Separable Transform
2D Wavelet Transform
Decomposition at Label 4
Original image(with diagonal details areas indicated)
Diagonal Details
Use Separable Transform
2D Wavelet Transform
Vertical Details
Decomposition at Label 4
Original image(with Vertical details areas indicated)
Experimental Results
Experimental Results
DWT
1.Original Image(Malignent_mdb238) 2.Decomposition at Label 4
2.Decomposition at Label 1 3.Decomposition at Label 2 3.Decomposition at Label 3
Experimental Results
DWT
1.Original Image(Malignent_mdb238) 2.Decomposition at Label 4
Experimental Results
1.Original Image(Benign_mdb252) 2.Decomposition at Label 4
2.Decomposition at Label 1 3.Decomposition at Label 2 3.Decomposition at Label 3
DWT
Experimental Results
1.Original Image(Malignent_mdb253.jpg) 2.Decomposition at Label 4
2.Decomposition at Label 1 3.Decomposition at Label 2 3.Decomposition at Label 3
Metrics: Quantitive Measurement
Metrics
To compare the ability of :Enhancement achieved by the proposed method
Why?
1. Measurement of distributed separation (MDS)2. Contrast enhancement of background against target (CEBT) and3. Entropy-based contrast enhancement of background against target (ECEBT) [23].
Measures used to compare:
23. Sameer S. and Keit B.: An Evaluation on Contrast Enhancement Techniques for Mammographic Breast Masses, IEEETransactions on Information Technology in Biomedicine, vol. 9, (2005) pp. 109-119
Metrics
1. Measurement of Distributed Separation (MDS)
Measures used to compare:
The MDS represents :How separated are the distributions of each mammogram
…………………………(3)MDS = |µucalcE -µtissueE |- |µucalc0 -µtissue0 |
µucalcE = Mean of the microcalcification region of the enhanced imageµucalc0 = Mean of the microcalcification region of the original image
µtissueE = Mean of the surrounding tissue of the enhanced imageµtissue0 = Mean of the surrounding tissue of the enhanced image
Defined by:
Where:
Metrics
2. Contrast enhancement of background against target (CEBT) Measures used to compare:
The CEBT Quantifies :The improvement in difference between the background and the target(MC).
…………………………(4)
0µucalcEµucalc
0µtissue0µucalc
EµtissueEµucalc
CEBT
σσ
−=
Defined by:
Where:
Eµucalcσ
0µucalcσ
= Standard deviations of the microcalcifications region in the enhanced image
= Standard deviations of the microcalcifications region in the original image
Metrics
3. Entropy-based contrast enhancement of background against target (ECEBT)Measures used to compare:
The ECEBT Measures :- An extension of the TBC metric- Based on the entropy of the regions rather
than in the standard deviations
Defined by:
Where:
…………………………(5)
0µucalcEµucalc
0µtissue0µucalc
EµtissueEµucalc
ECEBT
εζ
−=
= Entropy of the microcalcifications region in the enhanced image
= Entropy of the microcalcifications region in the original image
Eµucalcζ
0µucalcε
Experimental Results
MDS, CEBT and ECEBT metrics on the enhanced mammograms
Experimental Results
CT Method DWT Method
MDS CEBT ECEBT MDS CEBT ECEBT0.853 0.477 0.852 0.153 0.078 0.555
0.818 0.330 0.810 0.094 0.052 0.382
1.000 1.000 1.000 0.210 0.092 0.512
0.905 0.322 0.920 1.000 0.077 1.000
0.936 0.380 0.935 0.038 0.074 0.473
0.948 0.293 0.947 0.469 0.075 0.847
0.665 0.410 0.639 0.369 0.082 0.823
0.740 0.352 0.730 0.340 0.074 0.726
0.944 0.469 0.494 0.479 0.095 0.834
0.931 0.691 0.936 0.479 0.000 0.000
0.693 0.500 0.718 0.258 0.081 0.682
0.916 0.395 0.914 0.796 0.079 0.900
Table 1. Decomposition levels and directions.
0
0.2
0.4
0.6
0.8
1
1.2
TBC
Mammogram
MDS Matrix
CT DWT
The proposed method gives higher results than the wavelet-based method.
MDS, CEBT and ECEBT metrics on the enhanced mammograms
Experimental Results Analysis
0
0.2
0.4
0.6
0.8
1
1.2
TBCE
Mammogram
CEBT Matrix
CT DWT
The proposed method gives higher results than the wavelet-based method.
MDS, CEBT and ECEBT metrics on the enhanced mammograms
Experimental Results Analysis
0
0.2
0.4
0.6
0.8
1
1.2
DSM
Mammogram
ECEBT Matrix
CT DWT
The proposed method gives higher results than the wavelet-based method.
MDS, CEBT and ECEBT metrics on the enhanced mammograms
Experimental Results Analysis
Experimental Results AnalysisMesh plot of a ROI containing microcalcifications
(a)The original mammogram
(mdb252.bmp)
(b) The enhanced mammogram
using CT
Experimental Results Analysis
(a)The original mammogram
(mdb238.bmp)
(b) The enhanced mammogram
using CT
Experimental Results Analysis
(a)The original mammogram
(mdb253.bmp)
(b) The enhanced mammogram
using CT
More peaks corresponding to microcalcifications are enhanced
The background has a less magnitude with respect to the peaks:-The microcalcifications are more visible.
Observation:
Experimental Results Analysis
Experimental Results
(a)Original image (b)CT method (c)The DWT Method
These regions contain :• Clusters of microcalcifications (target)• surrounding tissue (background).
For visualization purposes :The ROI in the original mammogram are marked with a square.
ACHIEVEMENT
Improved Computer Assisted screen of mammogram
Achievements!
Enhancement of MC in digitized mammogramfor diagnostic support system
Figure: Diagnostic support system
MC
Suspected
Digital mammography systems :Presents images to the Radiologist with properly image processing applied.
Achievements!
(b) Enhanced image(mdb238.jpg)
(a) Original imageROI
(mdb238.jpg)
(a) Original imageWHOLE IMAGE (mdb238.jpg)
Digital mammography systems :Presents images to the Radiologist with properly image processing applied.
Hard to find MC Easy to find MC
Whilephysicians
interact with
The information in an image During interpretation process
Achievements!!
Enhancement of MC in digitized mammogram
With improved visual understanding, we can develop :
ways to further improve :o Decision making ando Provide better patient care
Improved Computer Assisted Screening
Goal Accomplished
Another Step Ahead..how about training a machine?
Dealing with Features
Why Feature Extraction?
Finding a feature:That has the most
discriminative information
The objective of feature selection:
Differs from its immediate surroundings by texture colorintensity
Fig: MC features (Extracted Using Human Visual Perception)
Why Feature Extraction?
Finding a feature:That has the most
discriminative information
The objective of feature selection:
Differs from its immediate surroundings by texture colorintensity
Fig: MC (Irregular in shape and size)(Extracted Using Human Visual Perception)
MoreFeatures: Shape Size
Why Feature Extraction?
Problems With MC Features:Irregular in shape and sizeNo definite patternLow Contrast -
Located in dense tissueHardly any color intensity variation
MC Feature
Fig: MC (Irregular in shape and size)(Extracted Using Human Visual Perception)
Why Feature Extraction? MC Feature
How radiologist deals with feature Detection/Recognition issue ?
Using Human Visual Perception
Why Feature Extraction? MC Feature
How Radiologist (Using Human Eye) deals with feature detection/Recognition issue ?
Using Human Visual Perception
Humans are equipped with sense organs e.g. eye-Eye receives sensory inputs and -Transmits sensory information to the brain
http://www.simplypsychology.org/perception-theories.html
Why Feature Extraction? MC Feature
Teach the machine to see like just we doObjective:
Irregular in shape and sizeNo definite patternLow Contrast -
Located in dense tissueHardly any color intensity variation
Machine Vision Challenges:-To make sense of what it sees
In Real:MC is Extracted Using Human Visual Perception
SURF Point Algorithm
Speeded-Up Robust Features (SURF) Algorithm
Point feature algorithm (SURF)Approach:
Improving the prediction performance of CAD Providing a faster, reliable and cost-effective prediction
Features will facilitate:
Fig: MC Point features (Extracted Using SURF point feature algorithm)
Point feature algorithm (SURF)Approach:
SURF point algorithm
Detect a specific object
Speeded-Up Robust Features (SURF) algorithm to find blob features.
Objective
based on Finding point correspondences
between . The reference and the target image
Reference Image Target Image
Feature Extraction
Context in using the features:I. Finding Key pointsII. Matching key pointsIII. Classification
Strongest feature point(Reference Image) Strongest feature point
(Target Image)
SURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
Feature Extraction
Strongest feature point(Reference Image)
Strongest feature point(Target Image)
SURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
Code Fragment (Detect and visualize feature points.)
%Detect feature points in the reference imageelephantPoints = detectSURFFeatures(elephantImage);
%Detect feature points in the target imagescenePoints = detectSURFFeatures(sceneImage);
% visualize feature points in the reference image.figure;imshow(elephantImage);hold on;plot(selectStrongest(elephantPoints, 100));title('100 Strongest Feature Points from Elephant Image');
% Extract Feature Points% Extract feature descriptors at the interest points in both images.
[elephantFeatures, elephantPoints] = extractFeatures(elephantImage, elephantPoints);[sceneFeatures, scenePoints] = extractFeatures(sceneImage, scenePoints);
Context in using the features:
Feature ExtractionSURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
I. Finding Key pointsII. Matching key pointsIII. Classification
Fig. Putatively Matched Points (Including Outliers )
Feature ExtractionSURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
elephantPairs = matchFeatures(elephantFeatures, sceneFeatures, 'MaxRatio', 0.9);
% Display putatively matched features.matchedElephantPoints = elephantPoints(elephantPairs(:, 1), :);matchedScenePoints = scenePoints(elephantPairs(:, 2), :);figure;showMatchedFeatures(elephantImage, sceneImage, matchedElephantPoints, ...
matchedScenePoints, 'montage');title('Putatively Matched Points (Including Outliers)');extractFeatures(sceneImage, scenePoints);
Code Fragment (Find Putative Point Matches)
Context in using the features:
Feature ExtractionSURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
I. Finding Key pointsII. Matching key pointsIII. Classification
Estimate Geometric Transformation and Eliminate Outliers
Context in using the features:
Feature ExtractionSURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
I. Finding Key pointsII. Matching key pointsIII. Classification
Feature ExtractionSURF point algorithmSpeeded-Up Robust Features (SURF) algorithm to find blob features.
% Estimate Geometric Transformation and Eliminate Outliers% estimateGeometricTransform calculates the transformation relating the matched points, % while eliminating outliers. This transformation allows us to localize the object in the scene[tform, inlierElephantPoints, inlierScenePoints] = ...
estimateGeometricTransform(matchedElephantPoints, matchedScenePoints, 'affine');figure;% Display the matching point pairs with the outliers removedshowMatchedFeatures(elephantImage, sceneImage, inlierElephantPoints, ...
inlierScenePoints, 'montage');title('Matched Points (Inliers Only)');
% Get the bounding polygon of the reference image.elephantPolygon = [1, 1;... % top-left
size(elephantImage, 2), 1;... % top-rightsize(elephantImage, 2), size(elephantImage, 1);... % bottom-right1, size(elephantImage, 1);... % bottom-left1,1]; % top-left again to close the polygon
newElephantPolygon = transformPointsForward(tform, elephantPolygon);
figure;imshow(sceneImage);hold on;line(newElephantPolygon(:, 1), newElephantPolygon(:, 2), 'Color', 'g');title('Detected Elephant');
CodeFragment
Moving Towards MC Feature DetectionUsing
SURF Point Algorithm
Local feature
Details In Next slide
To keep in mind
Local Feature Detection and Extraction
Local features :
A pattern or structure :Point, edge, or small image patch.
- A pattern or structure found in an image,
Differs from its immediate surroundings bytexture colorintensity
- Associated with an image patch that:
Fig.1 : Some Image Patch We used for Feature Point Detection Purpose
Local Feature Detection and Extraction
Applications: Image registration Object detection and classification TrackingMotion estimation
Using local features facilitates: handle scale changes rotation occlusion
Detectors /Methods :• FAST• Harris• Shi & Tomasi• MSER
• SURF
Feature Descriptors:SURFFREAKBRISKHOG descriptors
Detecting corner features
detecting blob/point features.
Speeded-Up Robust Features (SURF) algorithm to find blob features.
Detector Feature Type Scale IndependentFAST [24] Corner No
Minimum eigen value algorithm[25]
Corner No
Corner detector [26] Corner NoSURF [27] Blob/ Point YesBRISK [28] Corner YesMSER [29] Region with uniform
intensityYes
Local Feature Detection and Extraction
Why Using SURF Feature?Trying to identify MC cluster Blob
Speeded-Up Robust Features (SURF) algorithm to find blob features.
detectSURFFeatures(boxImage);
selectStrongest(boxPoints, 100)
extractFeatures(boxImage, boxPoints)
matchFeatures(boxFeatures, sceneFeatures);
Speeded-Up Robust Features (SURF) algorithm to find blob features.
Read the reference image containing the object of interest
Read the target image containing a cluttered scene.
Detect feature points in both images.
Select the strongest feature points found in the reference image.
Select the strongest feature points found in the target image.
Extract feature descriptors at the interest points in both images.
Find Putative Point Matches using their descriptors
Display putatively matched features.
Locate the Object in the Scene Using Putative Matches
Start
End
SURF Point Detection
1.Read the reference image
containing MC cluster
2.Target image containing MC.
2.Strongest feature point
in MC cluster
2. Strongest Feature point in Target Image
3. No match point Found
Speeded-Up Robust Features (SURF) algorithm to find blob features.
Are we getting less feature points?
Figure: No match point Found
No. of SURF feature points: 2 No. of SURF feature points: 47
Image Size256*256
Image Size 549*623
Image mdb238.jpg
More features from the image extracted(most points are mismatched)
To extract relevant feature point from the image
Case 1: Consider Big Reference Image
To get more feature points
To extract relevant feature point from the image
Case 2: Consider A bigger Reference Image andWhole mammogram as Target Image
1. Image of MC Cluster(mdb238.jpg) (256*256)
2. Main mammogram (mdb238.jpg) 1024*1024
3. 100 strongest point of ROI) (256*256) 4. 300 strongest point of Main mammogram (mdb238.jpg) 1024*1024
To get more feature points
What we finally have? No putative match Point
To extract relevant feature point from the image
Case 2: Consider A bigger Reference Image andWhole mammogram as Target Image To get more feature points
1. Image of an Microcalcification Cluster
Too small ROI will cause less feature points to match
2. 23 strongest pointsAmong 100 Strongest Feature Points
from reference image
Reference image: mdb248.jpgImage size: 256 *256
detectSURFFeatures(mc_cluster);
Problem 1: less number of feature points to matchSURF Feature Point
4. Only 1 strongest pointsAmong 300 Strongest Feature Points
from Scene Image
Too small ROI will cause less feature points to match
3. Image of a Cluttered Scene
Scene image: mdb248.jpgImage size: 427*588
detectSURFFeatures(sceneImage)
Problem 1: less number of feature points to matchSURF Feature Point
Result of small ROI (256*256):No Putative Point Matches
[mcFeatures, mc_Points] = extractFeatures(mc_cluster, mc_Points);[sceneFeatures, scenePoints] = extractFeatures(sceneImage, scenePoints);mcPairs = matchFeatures(mcFeatures, sceneFeatures);matchedmcPoints = mc_Points(mcPairs(:, 1), :);matchedScenePoints = scenePoints(mcPairs(:, 2), :);showMatchedFeatures(mc_cluster, sceneImage, matchedmcPoints, ... matchedScenePoints, 'montage');
Problem 1: less number of feature points to matchSURF Feature Point
Image Image Size Number of feature points
1190*589 15
588*427 23
256*256 1
541*520 86
Varying image size to see the effect to get SURF feature points
Approach-01 to solve: Considering the Whole image(Label and Pectoral Muscle)
Image size No. of SURF feature points
1024*1024 63
Target:To acquire more feature
2. Irrelevant Feature Points
Image size No. of SURF feature points
1024*1024 63
1. Less Feature points
Approach-01 to solve: Considering the Whole image(Label and Pectoral Muscle)
Target:To acquire more feature
Result:
Image size No. of SURF feature points
255*256 2
Approach-02 : Detect feature from the cropped image
Target:To acquire more feature
Image size No. of SURF feature points
256*256 2
Target:To acquire more feature
2. Relevant Feature Points
1. Less Feature pointsResult:
Approach-02 : Detect feature from the cropped image
Observation from approach 1 and 2
1. Image Size does not affectThe number of Feature Points
2. Zooming an image mayhelp to extract relevant featuresfrom the image(very few points to match)
mdb238.jpgImage Size: 1024*1024
mdb238.jpgImage Size: 256*256
Observation:Varying image size is not helping to get feature points
Image of an Microcalcification Cluster
23 strongest pointsAmong 100 Strongest Feature Points
from reference image
Reference image: mdb248.jpgImage size: 256 *256
Only 1 strongest pointsAmong 300 Strongest Feature Points
from Scene Image
Scene image: mdb248.jpgImage size: 427*588
Observing SURF Drawback
This method works best for :-- Detecting a specific object
(for example, the elephant in the reference image,rather than any elephant.)
-- Non-repeating texture patterns-- Unique feature
This technique is not likely to work well for:-- Uniformly-colored objects-- Objects containing repeating patterns.
detecting blob /point features. AIM Failed
Speeded-Up Robust Features (SURF) algorithm to find blob features.
Image Correlation Technique
Alternate ApproachImage Correlation Technique
Correlation
∑∑ ++=⊗k l
kjkihlkfhf ))((),(=f Image
=h Kernel/Mask
f1 f2 f3
f4 f5 f6
f7 f8 f9
h1 h2 h3
h4 h5 h6
h7 h8 h9
f1h1 f2h2 f3h3
f4h4 f5h5 f6h6
f7h7 f8h8 f9h9
=⊗ hf⊗
Experimental ResultsImage Correlation Technique
Image no: Benign mdb218.jpg
1. Original image
2. Kernel/ Mask/Template
3. Correlation Output
4. Identified MC(High value of sum.)
Image no: Benign mdb219.jpg
Image no: Benign mdb223.jpg
Image no: Benign mdb226.jpg
Image no: Benign mdb227.jpg
Image no: Benign mdb236.jpg
Image no: Benign mdb248.jpg
Image no: Benign mdb252.jpg
(Fixed Template Problem)..
Image no: Benign mdb222.jpg(Fixed Template Problem) Cont….
(Fixed Template Problem)..
Using Gabor Filter
Using Gabor Filter
• Make Gabor patch:
2; 2; 0.7854
2; 0.5; 0.7854
2; 2; 1.5708
5; 0.5; 1.5708
5; 2; 0.7854
2; 0.5; 1.5708
5; 0.5; 0.7854
5; 2; 1.5708
• Correlate the patch with image-To extract features of MC
⊗ =
0 10 20 30 40 50 60 70 80 90 100-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
Creating Gabor Mask
1. Linear RAMP
2. Linear RAMP values across: Columns Xm (left) and Rows Ym (Right)
3. Linear RAMP values across - Columns(Xm)
The result in the spatial domain
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5Xm (Across Columns) Ym- (Across rows)
4. Across Columns, Xm :a) Increase frequencyb )Use gray color map
6. Adding Xm and Ymtogether in
different proportions5. Across Rows, Ym :
a) Increase frequencyb )Use gray color map
Creating Gabor Mask
7. Create Gaussian Mask
8. Multiply Grating and Gaussian
Grating Gaussian Mask
Creating Gabor Mask
7. GABOR Mask
Creating Gabor Mask
Alternate ApproachUsing Gabor Filter Gabor kernel
2; 2; 0.7854
2; 0.5; 0.7854
2; 2; 1.5708
5; 0.5; 1.5708
5; 2; 0.7854
2; 0.5; 1.5708
5; 0.5; 0.7854
Scale , frequency, orientation
5; 2; 1.5708
MatrixSize = 26; %always scalar!
Scales = [2, 5];Orientations = [pi/4, pi/2];Frequencies = [0.5, 2];
CenterPoints = [13 13]; %int type (eg. [5 5; 13 13])
CreateMethod = FilterBank.CREATE_CROSSPRODUCT;
010
2030
020
40-0.5
0
0.5
2; 2; 0.7854
010
2030
020
40-0.2
0
0.2
2; 0.5; 0.7854
010
2030
020
40-0.2
0
0.2
2; 0.5; 1.5708
010
203
020
40-0.2
0
0.2
5; 2; 0.7854
010
2030
020
40-0.2
0
0.2
5; 2; 1.5708
010
2030
020
40-0.1
0
0.1
5; 0.5; 0.7854
010
2030
020
40-0.1
0
0.1
5; 0.5; 1.5708
010
2030
020
40-0.5
0
0.5
2; 2; 1.5708
Using Gabor Filter Gabor kernel
; 0 5; 5 08
Using Gabor Filter
⊗
⊗
⊗
=
=
=
Using Gabor Filter
⊗
⊗
⊗
=
=
=
⊗ =
Image In Spatial DomainUsing Gabor Filter Final Scenario
mini-MIAS drawbacksExperimental Realization
mini-MIAS drawbacksBenign mdb218
Original Enhanced
Gabor Effects
mini-MIAS drawbacksBenign mdb218
Original
Enhanced
2; 2; 0.7854
2; 0.5; 0.7854
2; 2; 1.5708
5; 0.5; 1.5708
5; 2; 0.7854
2; 0.5; 1.5708
5; 0.5; 0.7854
5; 2; 1.5708
Gabor Effects
mini-MIAS drawbacksBenign mdb218
Original
Enhanced
Gabor Effects
mini-MIAS drawbacksBenign mdb218
Original Enhanced
- NO definite Feature found
Observations:
Gabor Effects
mini-MIAS drawbacksBenign mdb218
Original Enhanced
Are these really enhanced?
-There is more detail, but could be noise.
-Enhanced versionseems to contain compression artifacts.
Question Arise?
Gabor Effects
mini-MIAS drawbacks
Enhanced version can contain Noise
Experimental Realization
1.Very Poor Quality with .jpg compression effects
a) Original image b) Enhanced image b) Enhanced imagea) Original image
mdb209
mdb213
mdb219
mdb249
mini-MIAS drawbacks
Not good enough for MC to be detectable
Experimental Realization
2. Reduced in resolution
Benign mdb218
Original Enhanced
mini-MIAS drawbacks
Not good enough for MC to be detectable
Experimental Realization
2. Reduced in resolution
Benign mdb218
Original
Enhanced
Where is MC?
OBSERVATION:
-There is more detail, but could be noise.
-Enhanced versionseems to contain compression artifacts.
More Evaluation (Gabor)mdb219.jpgBenign
OBSERVATION:
-Image Smoothing to remove edge will
Vanish the existenceof MC
More Evaluation (Gabor)mdb222.jpgBenign
OBSERVATION:
-NO definite feature of MC
More Evaluation (Gabor)mdb223.jpgBenign
OBSERVATION:
-NO definite feature of MCFalse contour
More Evaluation (Gabor)mdb223.jpgBenign
OBSERVATION:
-NO definite feature of MCFalse contour
No feature
More Evaluation (Gabor)mdb223.jpgBenign
OBSERVATION:
-NO definite feature of MCFalse contour
No feature
Several similar area false positive o/p
More Evaluation (Gabor)mdb226.jpgBenign
OBSERVATION:
- Bad resolution- Noise dominant- No definite feature of MC
More Evaluation (Gabor)mdb227.jpgBenign
OBSERVATION:
- Bad resolution/Poor quality image
- No definite feature of MC
More Evaluation (Gabor)mdb236.jpgBenign
OBSERVATION:
- Bad resolution-No definite feature of MC- Noise dominant
More Evaluation (Gabor)mdb240.jpgBenign
OBSERVATION:
- Bad resolution-No definite feature of MC- Noise dominant
More Evaluation (Gabor)mdb248.jpgBenign
OBSERVATION:
-feature of MC-But MC has different
orientationin different image
More Evaluation (Gabor)mdb252.jpgBenign
OBSERVATION:
-feature of MC-But MC has different
orientationin different image
More Evaluation (Gabor)mdb209.jpgMalignant
OBSERVATION:
- Bad resolution-No definite feature of MC- Noise dominant
More Evaluation (Gabor)mdb211.jpgMalignant
OBSERVATION:
- Bad resolution-No definite feature of MC- Noise dominant
More Evaluation (Gabor)mdb213.jpgMalignant
OBSERVATION:
- Bad resolution-No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb231.jpg
OBSERVATION:
- No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb238.jpg
OBSERVATION:
-No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb239.jpg
OBSERVATION:
-Image Smoothing to remove edge will
Vanish the existenceof MC
-No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb241.jpg
OBSERVATION:
-Image Smoothing to remove edge will
Vanish the existenceof MC
-No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb249.jpg
OBSERVATION:
-Image Smoothing to remove edge will
Vanish the existenceof MC
-No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb253.jpg
OBSERVATION:
- No definite feature of MC- Noise dominant
More Evaluation (Gabor)Malignant mdb256.jpg
Observation &
Drawing Conclusion
Future detection
Observation & Drawing Conclusion Feature Detection
• Reduced in resolution(Is not good enough for MC to be detectable)
• Very Poor Quality with .jpeg compression effects(Original MIAS doesn’t have such artifacts)
Limitations of mini-MIAS:
What can be done using mini-MIAS ?
• Can be used for big object detection(Pectoral Muscle, X-ray Label, Tumor, Mass detection)
Conclusion: mini-MIAS is not a good choice for:MC feature extraction
Observation & Drawing Conclusion Feature Detection
Any alternative to mini-MIAS?
Observation & Drawing Conclusion Feature Detection
Database Name Authority
MIAS ( Mammographic Image Analysis Society Digital Mammogram Database)
Mammography Image Analysis Society- an
organization of UK research groups
DDSM (Digital Database for Screening Mammogram) University Of South Florida
NDM (National Mammography Database) American College Of Radiology
LLNL/UCSF Database
Lawrence Livermore National Laboratories
(LLNL), University of California at San Fransisco (UCSF)
Radiology Dept.
Observation & Drawing Conclusion Feature Detection
Database Name Authority
Washington University Digital Mammography Database Department of Radiology at the
University of Washington
Nijmegen Database Department of Radiology at the
University of Nijmegen, the
Netherlands
Málaga mammographic database University of Malaga Central Research
Service (SCAI) ,Spain
BancoWeb LAPIMO Database Electrical Engineering Department at
Universidad de São Paulo, Brazil
Observation & Drawing Conclusion Feature Detection
These databases are NOT FREE
Research Findings
5; 0.5; 0.7854
Research FindingsImproved computer assisted
screening of mammogram
Detection and removal of big objects:- Pectoral Muscle - X-ray level
MC
Suspected
Observation & Drawing Conclusion On
Feature Detection
• Reduced in resolution(Is not good enough for MC to be detectable)
• Very Poor Quality with .jpeg compression effects(Original MIAS doesn’t have such artifacts)
Limitations of mini-MIAS:
What can be done using mini-MIAS ?
• Can be used for big object detection(Pectoral Muscle, X-ray Label, Tumor, Mass detection)
Conclusion: mini-MIAS is not a good choice for:MC feature extraction
BesideResearch Findings…
Published PaperAvailable Online:
http://cennser.org/IJCVSP/paper.html
Published PaperAvailable Online:
http://cennser.org/IJCVSP/paper.html
Published PaperAvailable Online:
http://cennser.org/IJCVSP/paper.html
Submitted Paperhttp://www.journals.elsevier.com/image-and-vision-computing/
1. Find Attribute/Feature From the enhanced mammogram:To train the machine:
-ANN (Artificial Neural Network)-SVM (Support Vector Machine)- GentleBoost Classifier [30]
2. Based on feature(size/shape), will move on to classification( benign or malignant)
MicrocalcificationIdentification
MicrocalcificationClassification
Plan of action as follows:
Further Research ScopeThere is always more to work on..In Research:
Future Plan
1. Segment the image
2. Find out the feature from the segmented image
3. Train the machine with features:
-ANN (Artificial Neural Network)-SVM (Support Vector Machine)
- GentleBoost Classifier [30]
4. Identify the MC5. Classify the MC
Available options
[2]D.Narain Ponraj, M.Evangelin Jenifer, P. Poongodi, J.Samuel Manoharan “A Survey on the Preprocessing Techniques of Mammogram for the Detection of Breast Cancer”, Journal ofEmerging Trends in Computing and Information Sciences, Volume 2, Issue 12, pp. 656-664,2011
[1]K. Santle Camilus , V. K. Govindan, P.S. Sathidevi,” Pectoral muscle identification inmammograms”, Journal of Applied Clinical Medical Physics , Vol. 12 , Issue No. 3 , 2011
[3]Arnau Oliver, Albert Torrent, Meritxell Tortajada, Xavier Llad´o,Marta Peracaula,Lidia Tortajada, Melcior Sent´ıs, and Jordi Freixenet,” Automatic microcalcification and clusterdetection for digital and digitised mammograms”, Springer-Verlag Berlin Heidelberg, 36,pp. 251–258, 2010
Reference
[4]Arnau Olivera, Albert Torrenta , Xavier Lladóa , Meritxell Tortajada, Lidia Tortajadab,Melcior Sentísb, Jordi Freixeneta, Reyer Zwiggelaarc,” Automatic microcalcification and clusterdetection for digital and digitised mammograms”,Elsevier:Knowledge-Based Systems,Volume 28, pp. 68–75, April 2012.
[5]A.Papadopoulos, D.I . Fotiadis, L.Costrrido,” Improvement of microcalcification clusterdetection in mammogaphy utilizing image enhancement techniques”.Comput.Bio.Med.10,Vol 38,Issue 38,pp.1045-1055,2008
[6]N.R.Pal,B.Bhowmik, S.K.Patel, S.Pal, J.Das,”A multi-stage nural network aided system fordetection of microcalcification in digitized mammogeams”,Neurocomputing, Vol 11,pp.2625-2634,2008
[7]M.Rizzi, M.D’Aloia, B.Castagnolo,” Computer aided detection of microcalcification in digitalMammograms adopting a wavelet decomposition ”,Integr.Comput.-Aided Eng.,Vol 16,Issue 2,pp.91-103,2009
Reference
[8]S.N.Yu, Y.K. Huang,” Detection of microcalcifications on digital mammograms using combinedModel-based and statistical textural features”, Expert Syst.Appl. , Vol 37,Issue 7,pp.5461-5469,2010
[9]Wang T. C and Karayiannis N. B.: Detection of Microcalci¯cations in Digital Mammograms Using Wavelets, IEEE Transaction on Medical Imaging, vol. 17, no. 4,(1989) pp. 498-509
[10]. Daubechies I.: Ten Lectures on Wavelets, Philadelphia, PA, SIAM, (1992)
[11] Strickland R.N. and Hahn H.: Wavelet transforms for detecting microcalcificationsin mammograms, IEEE Transactions on Medical Imaging, vol. 15, (1996) pp. 218-229
[12]Heinlein P., Drexl J. and Schneider Wilfried: Integrated Wavelets for Enhancement of Microcalcifications in Digital Mammography, IEEE Transactions on Medical Imaging, Vol. 22, (2003) pp. 402-413
[13]. Zhibo Lu, Tianzi Jiang, Guoen Hu, Xin Wang: Contourlet based mammographicimage enhancement, Proc. of SPIE, vol. 6534, (2007) pp. 65340M-1 - 65340M-8
Reference
[14]Fatemeh Moayedi, Zohreh Azimifar, Reza Boostani, and Serajodin Katebi: Contourlet-based mammography mass classification, ICIAR 2007, LNCS 4633,(2007) pp. 923-934
[15] Balakumaran T., Vennila ILA, Shankar C.G: Detection of Microcalcification in Mammograms Using Wavelet Transform and Fuzzy Shell Clustering, International Journal of Computer Science and Information Security, Vol 7,Issue 1,pp.121-125,2010
[16] Zhang X., Homma N., Goto S.,Kawasumi Y., Ihibashi T.,Abe M.,Sugita N.,Yoshizawa M: A Hybrid Image Filtering Method for Computer-Aided Detection of Microcalcification Clusters in Mammograms, Journal of Medical Engineering, Vol 3,Issue 1,pp.111-119,2013
[17] Lu J., Ikehara T., Zhang Y,Mihara T., Itoh T.,Maeda R:High quality factor silicon cantilever driven by piezoelectric thin film actuator for resonant based mass detection, Micro system Technologies , Vol 15, Issue 8, pp:1163-1169., 2009
[18] Leeuw H.D., Stehouwer BL, Bakker CJ, Klomp DW, Diest PV, Luijten PR, Seevinck PR,Bosch MA, Viergever MA, Veldhuis WB:Detecting breast microcalcifications with high-field MRI, NMR in Biomedicine,Vol 27, Issue 5, pages 539–546,2014
Reference
[19] Shankla V, David D. P, Susan P. Weinstein; Michael D., Tuite C, Roth R., Emily F:Automatic insertion of simulated microcalcification clusters in a software breast phantom, , Proc. SPIE 9033, Medical Imaging 2014: Physics of Medical Imaging, 2014
[20] Otsu, N., "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.
[21]Laine A.F., Schuler S., Fan J., Huda W.: Mammographic feature enhancement by multiscale analysis, IEEE Transactions on Medical Imaging, 1994, vol. 13, no. 4,(1994) pp. 7250-7260
22. Park S.-I., Smith M. J. T., and Mersereau R. M.: A new directional Filter bank forimage analysis and classification, Proceedings of IEEE International Conference onAcoustics, Speech, and Signal Processing (ICASSP '99), vol. 3, (1999) pp. 1417-1420
23. Sameer S. and Keit B.: An Evaluation on Contrast Enhancement Techniques forMammographic Breast Masses, IEEE Transactions on Information Technology inBiomedicine, vol. 9, (2005) pp. 109-119
Reference
24. Rosten, E., and T. Drummond. "Machine Learning for High-Speed Corner Detection." 9th European Conference on Computer Vision. Vol. 1, 2006, pp. 430–443.
25. Shi, J., and C. Tomasi. "Good Features to Track." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. June 1994, pp. 593–600.
26. Harris, C., and M. J. Stephens. "A Combined Corner and Edge Detector." Proceedings of the 4th Alvey Vision Conference. August 1988, pp. 147–152.
27 Bay, H., A. Ess, T. Tuytelaars, and L. Van Gool. "SURF: Speeded Up Robust Features." Computer Vision and Image Understanding (CVIU). Vol. 110, No. 3, 2008, pp. 346–359.
28.Leutenegger, S., M. Chli, and R. Siegwart. "BRISK: Binary Robust Invariant Scalable
29.Matas, J., O. Chum, M. Urba, and T. Pajdla. "Robust wide-baseline stereo from maximally stable extremal regions."Proceedings of British Machine Vision Conference. 2002, pp. 384–396.
Reference
30. Oliver A.; Torrent A. , Tortajada M, Liado X, R., Preacaula M , Tortajada L., Srntis M.,Ferixenet J: A Boosting based approach for automatic Microcalcification Detection,Springer-Verlag Berlin Heldelberg,Lecture notes on Computer Science (LNCS 6136), (2010)pp. 251- 258
Reference
Thank you foryour time and attention