Masters' whole work(big back-u_pslide)

Computer Assisted Screening of
Microcalcifications In Digitized
Mammogram For Early
Detection of Breast Cancer
Thesis Presentation
Nashid Alam
Registration No: 2012321028
annanya_cse@yahoo.co.uk
Supervisor: Prof. Dr. Mohammed Jahirul Islam
Department of Computer Science and Engineering
Shahjalal University of Science and TechnologyFriday, December 25, 2015
Driving research for better breast cancer treatment
“The best protection is early detection”
0
10
20
30
0
20
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
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
Mammography Machine

Mammography
USE:
I. Viewing x-ray image
II. Manipulate X-ray image on a computer screen
Mammography :
 Process of using low-energy
x-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
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

Masters' whole work(big back-u_pslide)

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 techniques
Boundary based method
Hybrid techniques
Watershed transformation
Edge detection:
Sobel
Prewitt
Roberts
Laplacian of Gaussian
Zero-cross
Canny
Goal:
oTo improve quality of image
oFacilate further processing
oRemove noise
oRemove 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 morphology
Local 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 algorithm
Redundant discrete wavelet linear stretching
and shrinkage algorithm.

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

Razzi et al.(2009)[7] proposed :
A two-stage decomposition wavelet filtering
First stage:
Reduce background noise
Second stage:
A hard thresholding technique:
-To identify microcalcification
Cluster was considered if more then 3 microcalcifications were
detected in a 1cm2 area
Literature Review

Yu et al.(2010)[8] :
Clustered microcalcification detection
used 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 region
Features 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 features
Emphasize 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.

Burdensome Task Of Radiologist :
Eye fatigue:
-Huge volume of images
-Detection accuracy rate tends to decrease
Non-systematic search patterns of humans
Performance gap between :
Specialized breast imagers and
general 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
Calcifications
Tumor
Lesion
Lump
Individual Research Areas
Problem Statement

Motivation to the research: Goal
Better Cancer Survival Rates
(Facilitate Early Detection ).
Provide “second opinion” : Computerized decision
support systems
Fast,
Reliable, and
Cost-effective
QUICKLY AND ACCURATELY :
Overcome the development of breast cancer

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 database
http://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
7
Normal
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 Semester
M-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 Semester
M-3

Schematic representation of the system

Schematicrepresentation
ofthesystem

Materials and Tools
Matlab 2014
Database: mini-MIAS

Removing Pectoral Muscle
And
X-ray Label

Image Segmentation
Goal: 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
(intensity <130) (intensity >200)
1
2
3
(a) Segmentation Part
(C) Final Segmented Image
(b)Original image
Why Segmentation?

Proposed framework for breast profile segmentation

Plan of Action:
1. Original Image 2. Segmentation Part
3. Final Segmented Image
4. Binary Image
Lactiferous Sinus, Ducts, lobule
(After removing pectoral muscles, fatty tissues, Ligaments)
(intensity <130) (intensity >200)
Separating the
Pectoral muscle
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.EROSION
II.DILATION
Using the basic operations we can perform -
a)OPENING
b)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
A
C
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 0
0 1 1 1 0
1 1 1 1 1
0 1 1 1 0
0 0 1 0 0
0 1 0
1 1 1
0 1 0

0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 1 0 0 0 0 0 0 0
0 0 1 1 1 1 1 0 0 0 0 0
0 1 1 1 1 1 1 1 0 0 0 0
0 1 1 1 1 1 1 1 0 0 0 0
0 0 1 1 1 1 1 1 0 0 0 0
0 0 1 1 1 1 1 1 1 0 0 0
0 0 1 1 1 1 1 1 1 1 1 0
0 0 0 0 0 1 1 1 1 1 1 0
0 0 0 0 0 0 0 0 0 0 0 0
B C
A
1 1 1
1 1 1
1 1 1
Structuring
Element 1
0 1 0
1 1 1
0 1 0
Structuring
Element 2

•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

Erosion of image f by structuring element s is
given by f  s
The structuring element s is positioned with its
origin at (x, y) and the new pixel value is
determined using the rule:
Erosion



=
otherwise0
fitsif1
),(
fs
yxg
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 removing
se = strel('disk',25);
for i=1:19
erode_bolb =
imerode(largest_bolb,se);
end
Original image
Erosion by
3*3
square
structuring
element
Erosion by
5*5
square
structuring
element
Watch out: Erosion shrinks objects

Erosion Example
Structuring Element
Original Image Processed Image With Eroded Pixels

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 s
The structuring element s is positioned with its
origin at (x, y) and the new pixel value is
determined using the rule:
⊕



=
otherwise0
hitsif1
),(
fs
yxg
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=1:9
BW2= imdilate (bw_image, [se1
se2], 'full')
BW2 = imfill(BW2,'holes');
end

Structuring Element
Original Image Processed Image
Dilation Example

Dilation Example
Structuring Element
Original Image Processed Image With Dilated Pixels

Dilation Example
Original image
Hole filling
Inside 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 bolb
se = strel('disk',39);
for i=1:19
closeBW_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 Analysis
OVER
-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
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
(intensity<150)(intensity>200)
1
2
3
(b) Segmentation Part
(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
(d)Binary Image
Achievement: X-Ray Label removed
Class: Benign

(b)Segmentation Part (c) Final Segmented Image(a)Main Image (e)Image containing
only Pectoral muscle
(d)Binary Image
What we need
mdb212.jpg
Issues 1.Biggest Cluster Does Not Contain Breast
Produce 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
(d)Binary Image
mdb214.jpg
mdb218.jpg
What We Need What We Have
Class: Benign

(e)Image containing duct, lobules,
sinus & Pectoral muscle
mdb001.jpg
(d)Binary Image
mdb222.jpg
mdb223jpg
mdb226jpg
Whatwewant
WhatweHave
Class: Benign

mdb001.jpg
(d)Binary Image
mdb240.jpg
mdb248.jpg
mdb252.jpg
Whatwewant
WhatweHave
(e)Image containing
duct, lobules, sinus & Pectoral musc
Class: Benign

Class: Malignant
mdb209 140 210
mdb211 140 210
mdb213 140 210
mdb216 140 210
mdb231 140 210
mdb233 140 210
mdb238 140 210
mdb239 140 210
mdb241 140 210
mdb245 140 210
mdb249 140 210
mdb253 140 210
mdb254 140 210
mdb256 140 210
(intensity<140)(intensity>210)
1
2
3
(b)Segmentation Part
(a)Original image
mdb209.jpg
(a) Original image

mdb241.jpg
(e)Image containing
duct, lobules, sinus
(d)Binary Image
Class: Malignant
Achievement: X-Ray Label removed
mdb238.jpg
mdb245.jpg
LabelRemoved

(a)Main Image
mdb212.jpg
Issues Level Remain In The Image
Class: Malignant
mdb209.jpg
(b)Segmentation Part
(c) Final Segmented Image
(d)Binary Image
(e)Image
Containing
only label

mdb212.jpg
Issues Pectoral muscle Remain In The Image
Produce Artifacts In Breast Region
Class: Malignant
(e)Image containing
only pectoral muscle
(d)Binary Image
mdb216.jpg
mdb213.jpg
(e)Image containing
only pectoral muscle
(d)Binary Image

Expected output: PECTORAL muscle, DUCT, LOBULES, SINUS,LIGAMENTS
mdb256.jpg
Output ImageMain image

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.
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

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

X-ray Label Removing Finding The Big BLOB
The types of noise :
High Intensity Rectangular Label
Low Intensity Label
Tape 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)

1.Binarizatin of original image.
2.Find the biggest blob.
Plan of Action:
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
Label successfully removed Issues

mdb212.jpg
mdb214.jpg
mdb218.jpg
mdb219.jpg
mdb222.jpg
mdb223.jpg
Class: Benign Issue with fatty tissues and ligaments existence

mdb226.jpg
mdb227.jpg
mdb236.jpg
mdb240.jpg
mdb248.jpg
mdb252.jpg
Class: Benign Issue with fatty tissues and ligaments existence

Original image
Binary Image
(threshold luminance level-=0.5) Original image Binary Image
Issue with fatty tissues and ligaments existenceClass: Malignant
mdb209.jpg
mdb211.jpg
mdb213.jpg
mdb216.jpg
mdb231.jpg
mdb233.jpg

Issue with fatty tissues and ligaments existenceClass: Malignant
mdb238.jpg
mdb239.jpg
mdb241.jpg
mdb245.jpg
mdb249.jpg
mdb253.jpg
mdb256.jpg

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;

Image Morphology
Experimental
results:
Goal: Region filling(Region inside the blob)
Original image
Finding 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

Experimental
results:
Original image
Result 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 Morphology

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 Morphology

-To find largest blob
Use -Otsu’s thresholding technique (graytrash) [20]
-Finding Bi-level the image(im2bw)
To Achieve The Desired Final Result:
-Apply
A 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.

1. Histogram equalization of the original X-ray image
2. Adjust image contrast
3. 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 region
6. Keep the true pixel value covering only the area of largest
blob and discard other features from the original image
7. 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.Histogram
Equalization
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 method
bw_image = im2bw(contrast_image, level);%getting binary image

5.Finding biggest blob
6.Hole filling
Inside 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

Experimental results

1.Original image
2.Histogram
Equalization 3.Contrast Image 4.Binary Image 5.Finding biggest blob
6.Hole filling
Inside the blob
7.Result image
(Label Removed)
mdb212.jpg
mdb214.jpg
mdb214.jpg
mdb218.jpg
mdb219.jpg
Benign

X-ray Label Removing Benign
1.Original image
2.Histogram
6.Hole filling
Inside the blob
7.Result image
(Label Removed)
mdb222.jpg
mdb223.jpg
mdb226jpg
mdb227jpg

X-ray Label Removing Benign
1.Original image
2.Histogram
6.Hole filling
Inside the blob
7.Result image
(Label Removed)
mdb226.jpg
mdb240.jpg
mdb248.jpg
mdb252.jpg

X-ray Label Removing Malignant
1.Original image
2.Histogram
6.Hole filling
Inside the blob
7.Result image
(Label Removed)
mdb209.jpg
mdb211.jpg
mdb213.jpg
mdb216.jpg
mdb231.jpg

1.Original image
2.Histogram
6.Hole filling
Inside the blob
7.Result image
(Label Removed)
mdb233.jpg
mdb238.jpg
mdb239.jpg
mdb241.jpg

1.Original image
2.Histogram
6.Hole filling
Inside the blob
7.Result image
(Label Removed)
mdb245.jpg
mdb249.jpg
mdb253.jpg
mdb256.jpg

Successful
Finally!

Removing pectoral muscle
Keeping fatty tissues and ligaments
mdb212.jpg
mdb213.jpg
(a)Main Image (b)Pectoral Muscle
mdb214.jpg
Main Image
Result Image

o Fatty tissue area
o Duct
o Lobules
o Sinus
o ligaments
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 muscle
Possible Approach To Edge-detection:
1.Scanning pixel value intensity at each points
2.find out the sudden big intensity change at the edge location
3.Mark the pixels at edge location
4.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 150 200
mdb214 130 205
mdb218 150 210
mdb219 120 200
mdb222 150 210
mdb223 150 225
mdb226 110 210
mdb227 150 230
mdb236 160 210
mdb240 150 200
mdb248 150 210
mdb252 140 210

Edge Detection of
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 muscle
1.Original image
mdb212.jpg
2.Segmentation Part
3.Fatty tissue
& Ligament removed
Possible Approach To Edge-detection(Approach-02):

Edge Detection of
pectoral muscle
4.Binary Version(outer)
5.Binary Version(inner)
6.Edge(outer-inner)

Edge Detection of
pectoral muscle

Edge Detection of
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 muscle
mdb223.jpg
mdb226.jpg
mdb240.jpg
mdb248.jpg
1.Original image 2.Segmentation Part
3.Fatty tissue
& Ligament removed4.Binary Version(outer)
5.Binary Version(inner) 6.Edge(outer-inner)

Edge Detection of
pectoral muscle
1.Pectoral muscle and ligaments in fatty tissue area got merged
mdb218.jpg
mdb240.jpg
3.Fatty tissue
Problems faced in (Approach-02):

Edge Detection of
pectoral muscle
2.Discontinuity in Pectoral muscle edge
mdb252.jpg
mdb226.jpg
mdb252.jpg
mdb248.jpg
3.Fatty tissue

Edge Detection of
pectoral muscle
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)
3.Fatty tissue
& Ligament removed
4.Binary Version(outer)
5.Binary Version(inner)
6.Edge(outer-inner)

Edge Detection of
pectoral muscle
Points to be noted from approach-2:
-Pectoral muscle a Triangular area
mdb212.jpg
mdb214.jpg
Based on this point:
Moving on to approach -03
mdb209.jpg
(2)Binary Image(1)Original Image

Triangle Detection
of pectoral muscle
1.Fing the triangular area of the pectoral muscle region
I. Finding white seeding point
II. Finding the 1st black point of 1st row after getting a white seeding point
III. Draw a horizontal line in these two points.
IV. finding the 1st black point of 1st column after getting a white seeding point
V. 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 muscle
mdb212.jpg
1.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 muscle
4.Triangle
5.Triangle Filled
6.muscle removed

Triangle Detection
of pectoral muscle

Triangle Detection
of 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 muscle
mdb222.jpg
mdb226.jpg
mdb227.jpg
2.Contrast stretching1.Original image 3.Binary of contrast image 4.Triangle
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 muscle
mdb218.jpg
1.Original image 2.Contrast stretching3.Binary of contrast image 4.Triangle
mdb219.jpg
mdb218.jpg
Class: Benign

Triangle Detection
of pectoral muscle
1.Original image 2.Contrast stretching3.Binary of contrast image 4.Triangle
mdb222.jpg
mdb219.jpg
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 muscle
Defects in mammogram
Class: Benign

2.Contrast stretching1.Original image
3.Binary of contrast image 4.Triangle
Triangle Detection
of pectoral muscle
Defects in mammogram
mdb227.jpg
Class: Benign

Triangle Detection
of 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 muscle
1.Original image
2.Contrast stretching
3.Binary of contrast image
4.Triangle
5.Triangle Filled
6.muscle removed
Class: Malignant
mdb256.pg

Successful
Pectoral Muscle Removing
Finally!

Improved
Computer Assisted Screening
Enhancement of digitized mammogram
Goal

MAIN NOVELTY
Input image
Bandpass
Directional
subbands
Bandpass
Directional
subbands

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.
Main Novelty

Enhancement of the Directional Subbands
The Contourlet Transform
Laplacian Pyramid: 3 level
Decomposition
Frequency 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

Decomposition
Laplacian Pyramid Level-1
8 Direction
4 Direction
4 Direction
(mdb252.jpg)

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

Decomposition
Contourlet coefficient at level 4
Wedge-shape frequency band is 23=8.
sub-bands 4-7
(5) sub-band 4
(6) sub-band 5
(7) sub-band 6
(8) sub-band 7

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

Directional 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

Directional filter banks
sub-bands 4-7
(5) sub-band 4
(6) sub-band 5
(7) sub-band 6
(8) sub-band 7

Why Contourlet?
•Decompose the mammographic image:
-Into directional components:
To easily capture the geometry of the image features.
Target

• 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 aspect
ratios
the aspect ratio of DWT is 1
CT similar as DWT can be
implemented using iterative filter banks.
Advantage of using 2D-CT over DWT:

Input image
Bandpass
Directional
subbands
Bandpass
Directional
subbands
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.

Method
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 and
classification, Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP '99), vol.
3, (1999) pp. 1417-1420
Details ……..

Method
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 and
classification, Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP '99), vol.
3, (1999) pp. 1417-1420

The CT is implemented by:
Laplacian pyramid followed by directional filter banks (Fig-01)
Input image
Bandpass
Directional
subbands
Bandpass
Directional
subbands
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
Bandpass
Directional
subbands
Bandpass
Directional
subbands
Details….
(c)
Denote
Each subband by yi,j
Where
i =decomposition level and
J=direction
Decomposes 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.
Denote
Where
J=direction Details….

Details….
The directional subbands are enhanced using equation 2.
=)( , jiyf
)2,1(,1 nnW jiy
)2,1(,2 nnW jiy
If bi,j(n1,n2)=0
If bi,j(n1,n2)=1
………..(2)
Denote
Where
J=direction
W1= weight factors for detecting the surrounding tissue
W2= 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

The directional subbands are enhanced using equation 2.
=)( , jiyf
)2,1(,1 nnW jiy
)2,1(,2 nnW jiy
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
Details….
The microcalcifications
appear :
On each subband
Over a very
homogeneous background.
Most of the transform coefficients:
-The coefficients corresponding to the
injuries 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
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.

Applying Contourlet Transformation Benign
Original image Enhanced image
Goal: Microcalcification Enhancement
mdb222.jpg
mdb223.jpg
mdb248.jpg
mdb252.jpg

mdb226.jpg
mdb227.jpg
mdb236.jpg
mdb240.jpg

Original image Enhanced image Original image Enhanced image
mdb218.jpgmdb219.jpg

Applying Contourlet Transformation Malignant
mdb209.jpg
mdb211.jpg
mdb213.jpg
mdb231.jpg

mdb238.jpg
mdb239.jpg
mdb241.jpg
mdb249.jpg

mdb253.jpg
mdb256.jpg

Applying Contourlet Transformation Normal
mdb003.jpg
mdb004.jpg
mdb006.jpg
mdb007.jpg

mdb009.jpg
mdb018.jpg
mdb027.jpg
mdb033.jpg

mdb046.jpg
mdb056.jpg
mdb060.jpg
mdb066.jpg

mdb070.jpg
mdb073.jpg
mdb074.jpg
mdb076.jpg

mdb093.jpg
mdb096.jpg
mdb101.jpg
mdb012.jpg

mdb128.jpg
mdb137.jpg
mdb146.jpg
mdb154.jpg

mdb166.jpg
mdb169.jpg
mdb224.jpg
mdb225.jpg

mdb263.jpg
mdb294.jpg
mdb316.jpg
mdb320.jpg

Use Separable Transform
2D Wavelet Transform
Visualization
Label of
approximation
Horizontal
Details
Horizontal
Details
Vertical
Details
Diagonal
Details
Vertical
Details
Diagonal
Details

Decomposition at
Label 4
Original image
(with diagonal details areas indicated)
Diagonal Details

Vertical Details
Decomposition at
Label 4
Original image
(with Vertical details areas indicated)

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

DWT
1.Original Image
(Malignent_mdb238) 2.Decomposition at Label 4

1.Original Image
(Benign_mdb252)
2.Decomposition at Label 4
DWT

1.Original Image
(Malignent_mdb253.jpg) 2.Decomposition at Label 4

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) and
3. 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, IEEE
Transactions on Information Technology in Biomedicine, vol. 9, (2005) pp. 109-119

Metrics
1. Measurement of Distributed Separation
(MDS)
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 image
µucalc0 = Mean of the microcalcification region of the original image
µtissueE = Mean of the surrounding tissue of the enhanced image
µtissue0 = Mean of the surrounding tissue of the enhanced image
Defined by:
Where:

Metrics
2. Contrast enhancement of background against
target (CEBT)
The CEBT Quantifies :
The improvement in difference between the background and the target(MC).
…………………………(4)
0µucalc
Eµucalc
0µtissue
0µucalc
Eµtissue
Eµucalc
CEBT
σ
σ
−
=
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)
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)
0µucalc
Eµucalc
0µtissue
0µucalc
Eµtissue
Eµucalc
ECEBT
ε
ζ
−
=
= Entropy of the microcalcifications region in the enhanced image
= Entropy of the microcalcifications region in the original image
Eµucalcζ
0µucalcε

MDS, CEBT and ECEBT metrics on the enhanced mammograms
CT Method DWT Method
MDS CEBT ECEBT MDS CEBT ECEBT
0.853 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.
Experimental Results Analysis

0
0.2
0.4
0.6
0.8
1
1.2
TBCE
Mammogram
CEBT Matrix
CT DWT

0
0.2
0.4
0.6
0.8
1
1.2
DSM
Mammogram
ECEBT Matrix
CT DWT

Mesh plot of a ROI containing microcalcifications
(a)The original
mammogram
(mdb252.bmp)
(b) The enhanced
mammogram
using CT

(a)The original
mammogram
(mdb238.bmp)
(b) The enhanced
mammogram
using CT

(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:

(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 mammogram
for 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 image
ROI
(mdb238.jpg)
(a) Original image
WHOLE IMAGE
(mdb238.jpg)
Digital mammography systems :
Presents images to the Radiologist
with properly image processing applied.
Hard to find MC Easy to find MC
While
physicians
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 and
o Provide better patient care
Improved
Computer Assisted Screening
Goal Accomplished

Another Step Ahead..how about training a machine?

Why Feature Extraction?
Finding a feature:
That has the most
discriminative information
The objective of feature selection:
Differs from its immediate surroundings by texture
 color
intensity
Fig: MC features (Extracted Using Human Visual Perception)

Finding a feature:
That has the most
discriminative information
The objective of feature selection:
Differs from its immediate surroundings by texture
 color
intensity
Fig: MC (Irregular in shape and size)
(Extracted Using Human Visual Perception)
More
Features:
 Shape
 Size

Problems With MC Features:
Irregular in shape and size
No definite pattern
Low Contrast -
Located in dense tissue
Hardly 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

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

Teach the machine to see like just we doObjective:
Irregular in shape and size
No definite pattern
Low Contrast -
Located in dense tissue
Hardly 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 points
II. Matching key points
III. Classification
Strongest feature point
(Reference Image)
(Target Image)

Feature Extraction
(Reference Image)
(Target Image)
Code Fragment (Detect and visualize feature points.)
%Detect feature points in the reference image
elephantPoints = detectSURFFeatures(elephantImage);
%Detect feature points in the target image
scenePoints = 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);

Feature ExtractionSURF point algorithm
III. Classification
Fig. Putatively Matched Points (Including Outliers )

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)

III. Classification
Estimate Geometric Transformation and Eliminate Outliers

III. Classification

% 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 removed
showMatchedFeatures(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-right
size(elephantImage, 2), size(elephantImage, 1);... % bottom-right
1, size(elephantImage, 1);... % bottom-left
1,1]; % 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');
Code
Fragment

Moving Towards MC Feature Detection
Using
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 by
texture
 color
intensity
- Associated with an image patch that:
Fig.1 : Some Image Patch We used for Feature Point Detection Purpose

Applications:
 Image registration
 Object detection and classification
 Tracking
 Motion estimation
Using local features
facilitates:
 handle scale changes
 rotation
 occlusion
Detectors /Methods :
• FAST
• Harris
• Shi & Tomasi
• MSER
• SURF
Feature Descriptors:
SURF
FREAK
BRISK
HOG descriptors
Detecting corner features
detecting blob/point features.

Detector Feature Type Scale Independent
FAST [24] Corner No
Minimum eigen value
algorithm[25]
Corner No
Corner detector [26] Corner No
SURF [27] Blob/ Point Yes
BRISK [28] Corner Yes
MSER [29] Region with uniform
intensity
Yes
Why Using SURF Feature?
Trying to identify MC cluster Blob

detectSURFFeatures(boxImage);
selectStrongest(boxPoints,100)
extractFeatures(boxImage,boxPoints)
matchFeatures(boxFeatures,sceneFeatures);
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

Are we getting less feature points?
Figure: No match point Found

No. of SURF feature points: 2 No. of SURF feature points: 47
Image Size
256*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

Case 2: Consider A bigger Reference Image and
Whole 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

What we finally have? No putative match Point
Case 2: Consider A bigger Reference Image and
Whole mammogram as Target Image

1. Image of an Microcalcification Cluster
Too small ROI will cause less feature points to match
2. 23 strongest points
Among 100 Strongest Feature Points
from reference image
Reference image: mdb248.jpg
Image size: 256 *256
detectSURFFeatures(mc_cluster);
Problem 1: less number of feature points to match
SURF Feature Point

4. Only 1 strongest points
from Scene Image
Too small ROI will cause less feature points to match
3. Image of a Cluttered Scene
Scene image: mdb248.jpg
Image size: 427*588
detectSURFFeatures(sceneImage)
SURF 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');
SURF 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
feature points
1024*1024 63
1. Less Feature points
Approach-01 to solve:
Considering the Whole image(Label and Pectoral Muscle)
Target:
Result:

feature points
255*256 2
Approach-02 :
Detect feature from the cropped image
Target:

feature points
256*256 2
Target:
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 affect
The number of Feature Points
2. Zooming an image may
help to extract relevant features
from the image
(very few points to match)
mdb238.jpg
Image Size: 1024*1024
mdb238.jpg
Image Size: 256*256

Observation:
Varying image size is not helping to get feature points
Image of an Microcalcification Cluster
23 strongest points
from reference image
Reference image: mdb248.jpg
Image size: 256 *256
Only 1 strongest points
from Scene Image
Scene image: mdb248.jpg
Image 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

Alternate Approach
Image 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
⊗

Image 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 mdb222.jpg
(Fixed Template Problem) Cont….

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
0.5
Xm (Across Columns)
Ym- (Across rows)

4. Across Columns, Xm :
a) Increase frequency
b )Use gray color map
6. Adding Xm and Ym
together in
different proportions
5. Across Rows, Ym :
a) Increase frequency
b )Use gray color map
Creating Gabor Mask

7. Create Gaussian Mask
8. Multiply Grating and Gaussian
GratingGaussian Mask
Creating Gabor Mask

7. GABOR Mask
Creating Gabor Mask

Alternate Approach
Using Gabor Filter Gabor kernel
2; 2; 0.78542; 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;

0
10
20
30
0
20
40
-0.5
0
0.5
2; 2; 0.7854
0
10
20
30
0
20
40
-0.2
0
0.2
2; 0.5; 0.7854
0
10
20
30
0
20
40
-0.2
0
0.2
2; 0.5; 1.5708
0
10
20
3
0
20
40
-0.2
0
0.2
5; 2; 0.7854
0
10
20
30
0
20
40
-0.2
0
0.2
5; 2; 1.5708
0
10
20
30
0
20
40
-0.1
0
0.1
5; 0.5; 0.7854
0
10
20
30
0
20
40
-0.1
0
0.1
5; 0.5; 1.5708
0
10
20
30
0
20
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 drawbacks
Experimental Realization

mini-MIAS drawbacks
Benign mdb218
Original Enhanced
Gabor Effects

mini-MIAS drawbacks
Benign mdb218
Original
Enhanced
2; 2; 0.78542; 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 drawbacks
Benign mdb218
Original
Enhanced
Gabor Effects

mini-MIAS drawbacks
Benign mdb218
Original Enhanced
- NO definite Feature found
Observations:
Gabor Effects

mini-MIAS drawbacks
Benign mdb218
Original Enhanced
Are these really enhanced?
-There is more detail,
but could be noise.
-Enhanced version
seems to contain
compression artifacts.
Question Arise?
Gabor Effects

mini-MIAS drawbacks
Enhanced version can contain Noise
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
2. Reduced in resolution
Benign mdb218
Original Enhanced

mini-MIAS drawbacks
Not good enough for MC to be detectable
2. Reduced in resolution
Benign mdb218
Original
Enhanced
Where is MC?
OBSERVATION:
-There is more detail,
but could be noise.
-Enhanced version
seems to contain
compression artifacts.

More Evaluation (Gabor)
mdb219.jpgBenign
OBSERVATION:
-Image Smoothing
to remove edge will
Vanish the existence
of MC

mdb222.jpgBenign
OBSERVATION:
-NO definite feature of MC

mdb223.jpgBenign
OBSERVATION:
False contour

mdb223.jpgBenign
OBSERVATION:
False contour
No feature

mdb223.jpgBenign
OBSERVATION:
False contour
No feature
Several similar area false positive o/p

mdb226.jpgBenign
OBSERVATION:
- Bad resolution
- Noise dominant
- No definite feature of MC

mdb227.jpgBenign
OBSERVATION:
- Bad resolution/Poor
quality image

mdb236.jpgBenign
OBSERVATION:
- Bad resolution
-No definite feature of MC
- Noise dominant

mdb240.jpgBenign
OBSERVATION:
- Bad resolution
- Noise dominant

mdb248.jpgBenign
OBSERVATION:
-feature of MC
-But MC has different
orientation
in different image

mdb252.jpgBenign
OBSERVATION:
-feature of MC
-But MC has different
orientation
in different image

mdb209.jpgMalignant
OBSERVATION:
- Bad resolution
- Noise dominant

mdb211.jpgMalignant
OBSERVATION:
- Bad resolution
- Noise dominant

mdb213.jpgMalignant
OBSERVATION:
- Bad resolution
- Noise dominant

Malignant mdb231.jpg
OBSERVATION:
- Noise dominant

OBSERVATION:
- Noise dominant

OBSERVATION:
-Image Smoothing
to remove edge will
of MC
- Noise dominant

OBSERVATION:
- Noise dominant

Observation
&
Drawing Conclusion
Future detection

Observation & Drawing Conclusion Feature Detection
• Very Poor Quality with .jpeg compression effects
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

Any alternative to mini-MIAS?

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.

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

These databases are NOT FREE

Research Findings
5; 0.5; 0.7854

Research Findings
Improved computer assisted
screening of mammogram
Detection and removal of big objects:
- Pectoral Muscle
- X-ray level
MC
Suspected

Observation & Drawing Conclusion On
Feature Detection
• Very Poor Quality with .jpeg compression effects
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
Beside
Research Findings…

Published Paper
Available Online:
http://cennser.org/IJCVSP/paper.html

Submitted Paper
http://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)
Microcalcification
Identification
Microcalcification
Classification
Plan of action as follows:
Further Research Scope
There 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 MC
5. 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 of
Emerging 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 in
mammograms”, 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 cluster
detection 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 cluster
detection 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 cluster
detection 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 for
detection 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 digital
Mammograms 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 combined
Model-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 microcalcifications
in 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 mammographic
image 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 for
image analysis and classification, Proceedings of IEEE International Conference on
Acoustics, Speech, and Signal Processing (ICASSP '99), vol. 3, (1999) pp. 1417-1420
23. Sameer S. and Keit B.: An Evaluation on Contrast Enhancement Techniques for
Mammographic Breast Masses, IEEE Transactions on Information Technology in
Biomedicine, 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 for
your time and attention

Masters' whole work(big back-u_pslide)

More Related Content

What's hot (20)

Similar to Masters' whole work(big back-u_pslide) (20)

Recently uploaded (20)

Masters' whole work(big back-u_pslide)