Slide 1 September 1999UCB The Analysis of Digital Mammograms: Spiculated Tumor Detection and Normal Mammogram Characterization Edward J. Delp Purdue University.

Slide 1September 1999 UCB

The Analysis of Digital Mammograms:Spiculated Tumor Detection

and Normal Mammogram Characterization

Edward J. Delp

Purdue University School of Electrical and Computer Engineering

Video and Image Processing Laboratory (VIPER)West Lafayette, Indiana, USA

[email protected]://www.ece.purdue.edu/~ace


Outline

• Breast Cancer and Mammography

• Multiresolution Detection of Spiculated Lesions

• Normal Mammogram Analysis and Characterization

• Future Research


Research Team

Charles Babbs - Department of Basic Medical Sciences

Zygmunt Pizlo - Department of Psychological Sciences

Sheng Lui - School of Electrical and Computer Engineering

Valerie Jackson - IU Department of Radiology

Funding - NSF, NIH, and Purdue Cancer Center

http://www.ece.purdue.edu/~ace/mammo/mammo.html


Breast Cancer

• Second major cause of cancer death among women in the United States (after lung cancer)

• Leading cause of nonpreventable cancer death

• 1 in 8 women will develop breast cancer in her lifetime

• 1 in 30 women will die from breast cancer


Mammography

• Mammograms are X-ray images of the breast

• Screening mammography is currently the best technique for reliable detection of early, non-palpable, potentially curable breast cancer

• Studies show that mammogram can reduce the overall mortality from breast cancer by up to 30%


Screening Mammography


A Digital Mammogram (normal)


Three Types of Breast Abnormalities

Micro-calcification

Circumscribed Lesion

Spiculated Lesion


Problems in Screening Mammography

• Radiologists vary in their interpretation of the same mammogram

• False negative rate is 4 – 20% in current clinical mammography

• Only 15 – 34% of women who are sent for a biopsy actually have cancer


Current Research in Computer Aided Diagnosis (CAD)

• The goal is to increase diagnostic accuracy as well as the reproducibility of mammographic interpretation

• Most work aims at detecting one of the three abnormal structures

• Some have explored classifying breast lesions as benign or malignant

• The implementation of CAD systems in everyday clinical applications will change the practice of radiology


Multiresolution Detection of Spiculated Lesions in Digital Mammograms

• Spiculation or a stellate appearance in mammograms indicates with near certainty the presence of breast cancer

• Detection of spiculated lesions is very important in the characterization of breast cancer


Spiculated Lesions

• Spiculated lesions vary from a few millimeters to several centimeters in size

• Center masses of spiculated lesions are usually irregular with ill-defined borders

• Usually the larger the tumor center, the longer its spicules or “arms”


Difficulties

• Computer aided diagnosis of digital mammograms generally consists of feature extraction followed by classification

• It is very difficult to determine the neighborhood size that should be used to extract features which are local

• If the neighborhood is too large, small lesions may be missed

• If the neighborhood is too small, one may not be able to capture features of larger lesions


Appearance of A Spiculated Lesion at Multiple Resolutions


Block Diagram of Multiresolution Detection of Spiculated Lesions


Multiresolution Decomposition

• Linear phase nonseparable 2D perfect reconstruction wavelet transform

– does not introduce phase distortions in the decomposed images

– no bias is introduced in the horizontal and vertical directions as a separable transform would

• The impulse response of the analysis low pass filter

0125.00

125.05.0125.0

0125.00

)n,n(h 21


A Spiculated Lesion Distorts the Normal Breast Duct Structure

• Normal duct structures of the breast radiate from the nipple to the chest wall

• Spiculated lesion radiates spicules in all directions


Gradient Orientation Histogram

• Has a peak at the ductal structure orientation near a normal pixel

• Flat near a lesion pixel


Example Histograms

A normal region

A spiculated region


Notation

• (i, j) — spatial location at row i and column j

• f(i, j) — pixel intensity at (i, j)

Sij — some neighborhood of (i, j)

• M — the number of pixels within Sij

• Dy(i, j) and Dx(i, j) — estimate of the vertical and horizontal spatial derivatives of f at (i, j), respectively

(i, j) = tan-1{Dy(i, j)/Dx(i, j)} (-/2, /2] — estimate of the gradient orientation at (i, j)


Notation

• histij — histogram of within Sij using 256 bins

• histij(n) — # of pixels in Sij that have gradient

orientations , where n = 0, 1,

…, 255

• — average bin height of histij

255

0nij )n(hist

2561

)j,i(hist

256)1n(

2,

256n

2


Folded Gradient Orientation

• M+(i, j) and M-

(i, j) — number of positive and negative

gradient orientations within Sij, respectively

• and — average positive and negative

gradient orientations, respectively

•

— folded gradient orientation

)j,i( )j,i(

otherwise)j,i(

)j,i(M)j,i(M,2/)j,i()j,i(if)j,i(

)j,i(M)j,i(M,2/)j,i()j,i(if)j,i(

)j,i(


Why Folded Gradient Orientation?

So that is not sensitive to the nominal value of , but to the actual gradient orientation variances

)j,i(

• The gradient orientation distance between /2 and -/4 is the same as that between /2 and /4, however ([/2, -/4]) = 2.8 ([/2, /4]) = 0.3

• -/4 folds to 3/4, now

’([/2, -/4]) = ’([/2, /4]) = 0.3


Features Differentiate Spiculated Lesions from Normal Tissue

ijS)n,m(

)n,m(fM1

)j,i(f

• Mean pixel intensity in Sij —

• Standard deviation of pixel intensities in Sij —

ijS)n,m(

2f ))j,i(f)n,m(f(

1M

1)j,i(


Features Differentiate Spiculated Lesions from Normal Tissue

• Standard deviation of gradient orientation histogram in Sij —

• Standard deviation of the folded gradient orientations in Sij —

255

0n

2ijijhist ))n(hist)n(hist(

255

1)j,i(

ijS)n,m(

2))j,i()n,m((1M

1)j,i(


Multiresolution Feature Analysis

• Choose a neighborhood that is small enough to capture the smallest possible spiculated lesion in the finest resolution

• Fix this neighborhood size for feature extraction at all resolutions

• Larger lesions will be detected at a coarser resolution

• Smaller lesions can be detected at a finer resolution


Test Pattern at Multiple Resolutions

An ideal spiculated lesion and normal duct structures embedded in uncorrelated Gaussian distributed noise


Feature ’ at Multiple Resolutions


Feature hist at Multiple Resolutions


Feature at Multiple Resolutionsf


Feature f at Multiple Resolutions


A Simple Binary Tree Classifier


Advantages ofTree-Structured Approach

• Robust with respect to outliers and misclassified points in the training set

• The classifier can be efficiently represented

• Once trained, classification is very fast

• Provides easily understood and interpreted information regarding the predictive structure of the data

• Classifier used is described in a paper by Gelfand, Ravishankar, and Delp (PAMI 1991)


Multiresolution Detection

• At each resolution, five features are used: the four features extracted at that resolution plus the feature hist extracted from the next coarser resolution

• Detection starts from the second coarsest resolution

• A positive detection at a coarser resolution eliminates the need for both feature extraction and detection at the corresponding pixel locations at all finer resolutions

• A negative result at a coarser resolution will be combined with those at finer resolutions via weighted sum


Database

• MIAS database provided by the Mammographic Image Analysis Society in the UK

• 50 resolution

• A total of 19 mammograms containing spiculated lesions

• Smallest lesion extends 3.6 mm in radius

• Largest lesion extends 35 mm in radius


Half/Half Training Methodology

• The 19 mammograms containing spiculated lesions together with another 19 normal mammograms are random split into two sets with approximately an equal number of lesion and normal mammograms in each set

• Each set was used separately as a training set to generate two BCTs

• A BCT trained by one set was used to classify mammograms in the other set, and vice versa


Detection Results

A 35.0mm lesion detected at the coarsest resolution

Automatic Detection Ground Truth


Detection Results


A 12.4mm lesion detected at the second coarsest resolution


Detection Results

A 6.6mm lesion detected at the finest resolution



Summary

• Multiresolution detection eliminates the problem of choosing a neighborhood size a priori to capture features of lesions of varying sizes

• Using features across resolutions simultaneously helps capture spiculated lesions of sizes that exist between the resolutions

• Top-down approach requires less computation by starting with the least amount of data and propagating detection results to finer resolutions


Normal Mammograms Characterization

• Better understanding of normal mammograms can greatly help reduce the “misses” in cancer detection

• Little work has been done on characterizing normal mammograms

P(normal)normal)|P(imageP(cancer)cancer)|P(image

P(cancer)cancer)|P(imageimage)|P(cancer


Very Different Normal Mammograms

Density 1 Density 2 Density 3 Density 4


General Normal Characteristics

• Unequivocally normal areas have lower density than abnormal ones

– no spikes indicating microcalcifications

– no large bright areas indicating masses

• Normal areas have “quasi-parallel” linear markings


Normal Linear Markings

• Shadow of normal ducts and connective tissue elements

• Appear slightly curved

• Approximately linear over short segments

• Can be observed as straight line segments of dimensions 1 to 2 mm or greater in length and 0.1 to 1.0 mm in width

• Low contrast in very noisy background


Recognizing Normal Structures


Background Subtraction


Problems in Extracting Linear Markings

• Edge detection based line detectors

– generate very dense edge maps due to small spatial extent of most local edge operators

– miss “thick” lines

• Hough transform based line detectors

– do not provide locations of lines

– not suitable for grayscale images


A New Model For Lines

• There exists a string of pixels with similar graylevels along a certain direction

• The surrounding pixels have different graylevels

• The length of a line is greater than its width


Line Detection Block Diagram


Advantages

Given minimum length l, our new line detector can detect

– lines of very different width, from single pixel wide up to l

– lines of any length that is greater than l

– lines with varying width, provided that the changes are “slower” than l

– curves, provided that over short segment, they can be approximated as lines of length greater than l


Detect Lines in Noisy Image


Detect Lines Narrower Than l

Test Image l = 5 l = 20


Normal Structure Removal


Database

• Digital Database for Screening Mammography (DDSM) provided by Massachusetts General Hospital, University of South Florida, and Sandia National Laboratories

• 42 / 50 • More than 650 cases available now

• Each case consists of 4 images: left and right MLO and CC views

• Have pixel level “ground truth” information


Residual Image of Circumscribed Lesions

Original Mammogram

Background Subtracted Image

Residual Image


Residual Image of Spiculated Lesion

Original Mammogram


Residual Image


Residual Image of Microcalcifications

Original Mammogram


Residual Image


Compare with Strickland’s Microcalcification Enhancement Algorithm

Original Mammogram

Strickland’s Enhanced Image

Residual Image


Compare with Strickland’s Microcalcification Enhancement Algorithm

Original Mammogram

Strickland’s Enhanced Image

Residual Image


Summary

• Normal mammogram characterization is fundamentally simpler — characteristics of normal tissue are relatively simpler than characteristics of tumors of various types, sizes, and stages of development

• Suppressing normal structures essentially enhances abnormal structures — facilitates the classification of abnormalities


Future Research

• Detect stellate patterns in the “Normal” mask before removal so that spiculation remains in the residual image

• Detect normal mammogram based on obtained residual images

Slide 1 September 1999UCB The Analysis of Digital Mammograms: Spiculated Tumor Detection and Normal Mammogram Characterization Edward J. Delp Purdue University.

Documents