A Method for Producing Simulated Mammograms: A Method for Producing Simulated Mammograms: Observer StudyObserver Study

Payam Seifi M.Sc.Payam Seifi M.Sc.

Michael R. Chinander Ph.D.Michael R. Chinander Ph.D.

Robert M. Nishikawa Ph.D., FAAPMRobert M. Nishikawa Ph.D., FAAPM

Carl J. Vyborny Translational Laboratory for Breast Imaging ReseCarl J. Vyborny Translational Laboratory for Breast Imaging Researcharch

Department of RadiologyDepartment of Radiology

Committee on Medical PhysicsCommittee on Medical Physics

The University of ChicagoThe University of Chicago

IntroductionIntroduction

Large databases with known truth are needed to Large databases with known truth are needed to develop CAD systemsdevelop CAD systemsLimitations of databases:Limitations of databases:Accurate truth information is not available Accurate truth information is not available

exact mass borders, exact shape and size of exact mass borders, exact shape and size of microcalcificationsmicrocalcifications

Optimization of CAD depends on imaging system Optimization of CAD depends on imaging system propertiesproperties

transition to a new imaging system requires new transition to a new imaging system requires new dataset, and acquiring truth for all lesionsdataset, and acquiring truth for all lesions

OverviewOverview

Developed a method for producing simulated Developed a method for producing simulated mammogaphic imagesmammogaphic images

Performed observer study to evaluate the quality Performed observer study to evaluate the quality of the simulated imagesof the simulated images

Part I: Simulating MammogramsPart I: Simulating Mammograms

The method proposed here simulates The method proposed here simulates mammograms with accurate truth information:mammograms with accurate truth information:

HighHigh--quality radiographs of cadaver breasts, quality radiographs of cadaver breasts, mastectomy, and biopsy specimensmastectomy, and biopsy specimensCombining the radiographs of the cadaver breast with Combining the radiographs of the cadaver breast with lesion radiographslesion radiographsThe combined image used as input to the model of the The combined image used as input to the model of the image formationimage formationAccurate model of the process of image formation in a Accurate model of the process of image formation in a mammographic screenmammographic screen--film systemfilm system

1. Input Images1. Input Images

High fidelity images of cadaver breast specimensHigh fidelity images of cadaver breast specimensFaxitron MXFaxitron MX--20 system20 systemdirect exposure Xdirect exposure X--OmatOmat TL filmTL filmgeometric magnification (2X) geometric magnification (2X) obtained high resolution and low noise radiographsobtained high resolution and low noise radiographs

Digitized at 50Digitized at 50μμm pixel size m pixel size 2525μμm in the object planem in the object planeLumiscanLumiscan 85 laser film scanner85 laser film scanner

Input ImageInput Image2.

5 cm

Input Image

Pixel value tox-ray quanta conversion

Add Poisson

noise

Add Scatter

Simulation ofphosphor

screen

Optical density,

filmnoise

Light quanta topixel valueconversion

OutputImage

Flowchart of the Components of the Flowchart of the Components of the SimulationSimulation

φμ

νρ

⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠

0.00873 J=X abh kg.R

air

2. X2. X--ray Quanta Distributionray Quanta Distribution

We calibrated the system by:We calibrated the system by:

Acquiring images with different thicknesses of Lucite and Acquiring images with different thicknesses of Lucite and exposure times using direct xexposure times using direct x--ray filmray film

Measuring the exposure for each thickness with a dosimeter Measuring the exposure for each thickness with a dosimeter (model 10100A, Keithley Radiation Masurement),(model 10100A, Keithley Radiation Masurement),

The photon fluence The photon fluence ΦΦ per unit exposure X was estimated per unit exposure X was estimated by:by:

2. X2. X--ray Quanta Distributionray Quanta Distribution

The xThe x--ray quanta are assumed to be monoray quanta are assumed to be mono--energetic with 20 energetic with 20 keVkeV energyenergy

Calculated number of photons per pixel of the Calculated number of photons per pixel of the high fidelity imagehigh fidelity image

Noise was added by a random number generatorNoise was added by a random number generator

Poisson distribution with a mean equal to the number Poisson distribution with a mean equal to the number of photons per pixelof photons per pixel

3. Scattered Radiation3. Scattered Radiation

Low scatter in the input images, due to Low scatter in the input images, due to magnificationmagnification

Compensated by estimating the scatter field Compensated by estimating the scatter field using the simulation data (Boone using the simulation data (Boone et alet al. 2000). 2000)

Simulation data from 4cm thick breast and 28 kVpSimulation data from 4cm thick breast and 28 kVpxx--ray beamray beam

A symmetric PSF was generated and convolved A symmetric PSF was generated and convolved with the input data to estimate the scatter fieldwith the input data to estimate the scatter field

Example: xExample: x--ray field and scatter fieldray field and scatter field

Input image Scatter image

3. Model of the Phosphor Screen3. Model of the Phosphor Screen

Depth dependent MTF was used to account for Depth dependent MTF was used to account for the thickness of the scintillating screenthe thickness of the scintillating screen

For each depth, MTF is obtained by diffusion For each depth, MTF is obtained by diffusion equation approximation (Swank 1973) equation approximation (Swank 1973)

Parameters used in the simulation were Parameters used in the simulation were determined by matching the MTF and NPS to determined by matching the MTF and NPS to experimental data from Minexperimental data from Min--R 2000 screen R 2000 screen (courtesy P. Bunch, Eastman Kodak company)(courtesy P. Bunch, Eastman Kodak company)

Phosphor ScreenPhosphor Screen

X-ray

Phosphor

Film

Screen MTFScreen MTF

experimentX simulated

Screen NPSScreen NPS

The light photon density onto the film due to xThe light photon density onto the film due to x--ray ray interaction at depth t is:interaction at depth t is:

r(x,y,tr(x,y,t) output light quanta ) output light quanta fluencefluence, , q(x,y,tq(x,y,t) input x) input x--ray ray fluencefluence, , ΦΦ((u,v,tu,v,t)): OTF at depth t.: OTF at depth t.

The resulting distribution from each sublayer is summed The resulting distribution from each sublayer is summed to create the final distribution of light quantato create the final distribution of light quanta

The light distribution output was converted to optical The light distribution output was converted to optical density using H&D curve of the filmdensity using H&D curve of the film

)}},,(),,({{),,( ,1 tvutyxqFTFTtyxr vuΦ= −

−= Φ1,( , , ) { { ( , , )} ( , , )}u vr x y t FT FT q x y t u v t

3. Model of the Phosphor Screen3. Model of the Phosphor Screen

Example: Input and output images of the Example: Input and output images of the phosphor screenphosphor screen

Input x-ray distribution onto the phosphor

Output light distribution from the phosphor

4. Film Granularity Noise4. Film Granularity Noise

Gaussian white noise was generated for each Gaussian white noise was generated for each model pixel filtered by the shape of the film NPS model pixel filtered by the shape of the film NPS for Minfor Min--R 2000, and added to screen output image R 2000, and added to screen output image

The characteristic curve of the digitizer was used The characteristic curve of the digitizer was used to convert from optical density to pixel valueto convert from optical density to pixel value

Real and Simulated ImagesReal and Simulated Images

Screen-film image Simulated image

Part II: Observer StudyPart II: Observer Study

The goal of the study is to determine how well The goal of the study is to determine how well observers can distinguish simulated mamograms observers can distinguish simulated mamograms from real onesfrom real onesThe quality of the simulated mammograms is The quality of the simulated mammograms is evaluated using an observer studyevaluated using an observer studyThe ROC curves are generated for each observerThe ROC curves are generated for each observerThe area under the ROC curve (AUC) is computedThe area under the ROC curve (AUC) is computed

AUC indicates the accuracy with which they can AUC indicates the accuracy with which they can distinguish the two types of imagesdistinguish the two types of images

Observer Program InterfaceExample Simulated Image Example Screen/film Image

0

100

Overview of the Observer ProgramOverview of the Observer Program

Six training images are givenSix training images are givenThe observer indicates their confidence that the The observer indicates their confidence that the image is a real mammogram on a continuous 0image is a real mammogram on a continuous 0--100 scale100 scaleEach observer reads 130 images (half simulated Each observer reads 130 images (half simulated and half real)and half real)

Observer StudyObserver Study

For the preliminary study, five nonFor the preliminary study, five non--radiologist radiologist observers were asked to complete the taskobservers were asked to complete the task

All were experienced (> 4 yrs) in CAD for All were experienced (> 4 yrs) in CAD for mammography and were familiar with the appearance mammography and were familiar with the appearance of mammogramsof mammograms

We used the LABROC program to generate ROC We used the LABROC program to generate ROC curves and calculate the AUC for each observercurves and calculate the AUC for each observer

Results: ROC curvesResults: ROC curves

Results: AUC ValuesResults: AUC Values

AverageAverage

55

44

33

22

11

ObserverObserver

0.042*0.042*0.5300.530

0.0500.0500.5380.538

0.0500.0500.5270.527

0.0500.0500.5510.551

0.0490.0490.50.57272

0.0490.0490.4610.461

Standard Standard ErrorError

AUCAUC(LabROC)(LabROC)

* Standard Deviation of AUC values

ConclusionsConclusions

The observer study showed that observers could The observer study showed that observers could not distinguish simulation images from real not distinguish simulation images from real screenscreen--film images film images Our method can produce realistic simulated Our method can produce realistic simulated mammogramsmammograms

Future WorkFuture Work

We will repeat the observer study using expert We will repeat the observer study using expert breast radiologistsbreast radiologistsWe will use our method for:We will use our method for:

developing CAD systemsdeveloping CAD systemsdeveloping and optimizing xdeveloping and optimizing x--ray imaging systemsray imaging systems