Breast Cancer - FACTS: Breast carcinoma leading cause of cancer death in womean Every 8-10 th woman affected during lifetime About 4000 new cases/a.

Breast Cancer - FACTS:Breast Cancer - FACTS:Breast Cancer - FACTS:Breast Cancer - FACTS: Breast carcinoma leading Breast carcinoma leading

cause of cancer death in cause of cancer death in womeanwomean

Every 8-10Every 8-10thth woman woman affected affected during lifetimeduring lifetime

About 4000 new cases/a About 4000 new cases/a in in Austria Austria

Clustered Clustered mircrocalcifications one of mircrocalcifications one of early sign‘searly sign‘s

Breast Cancer - FACTS:Breast Cancer - FACTS:Breast Cancer - FACTS:Breast Cancer - FACTS:

0

1020

30

4050

6070

80

90100

Total local regional Metastasis unstaged

SURVIVAL vs TUMORSPREAD

Mammography - Mammography - TECHNIQUE:TECHNIQUE:

Mammography - Mammography - TECHNIQUE:TECHNIQUE:

Mammography - Mammography - REPORTING:REPORTING:

Mammography - Mammography - REPORTING:REPORTING:

Clsutered Microcalcifications are Clsutered Microcalcifications are early sign´s ofbreast cancer early sign´s ofbreast cancer (bright spot´s)(bright spot´s)- Size: 0.2-0.5 mm - search with a Size: 0.2-0.5 mm - search with a

magnifying glassmagnifying glass- Difficult perception in low contrast areasDifficult perception in low contrast areas- Differentiate: cancerous vs non Differentiate: cancerous vs non

cancerouscancerous

Mammography -Mammography -DILEMMA: DILEMMA:

Mammography -Mammography -DILEMMA: DILEMMA:

Special training of radiologists Special training of radiologists necessary necessary

Positive predictive value of Positive predictive value of radiologists: 20%radiologists: 20%

„ „Double Reading“ (independet Double Reading“ (independet reporting by 2 reporting by 2 radiologists): radiologists): improves accuracy by 5 - 15%improves accuracy by 5 - 15%

Artificial Intelligence in Artificial Intelligence in Mammography - Mammography -

HYPOTHESIS:HYPOTHESIS:

Artificial Intelligence in Artificial Intelligence in Mammography - Mammography -

HYPOTHESIS:HYPOTHESIS:

Clustered microcalcifications Clustered microcalcifications be found reliably found by a be found reliably found by a computer application computer application

and a discrimination between and a discrimination between cancerous and non cancerous cancerous and non cancerous microcalcifications be done microcalcifications be done using neuronal netsusing neuronal nets

Mammography - Mammography - VISIONVISION

Mammography - Mammography - VISIONVISION

CAD System CAD System ((CComputer omputer AAided ided DDiagnosis) can act as iagnosis) can act as “never tired, second “never tired, second reader”reader”

Variance of female Variance of female FEATURESFEATURES

Variance of female Variance of female FEATURESFEATURES

ANN & Mammography - ANN & Mammography - PATIENTS: PATIENTS:

ANN & Mammography - ANN & Mammography - PATIENTS: PATIENTS:

Patient Database:Patient Database:- 100 patients = 272 Images 100 patients = 272 Images

with Mc´swith Mc´s- High resolution film High resolution film

digitalization: digitalization: 8000x6000x15 -> about 90 8000x6000x15 -> about 90 Mbyte/imageMbyte/image

- Resolution for image Resolution for image processing: 91.5 processing: 91.5 mm

ANN & Mammography - ANN & Mammography - GROUNDTRUTH:GROUNDTRUTH:

ANN & Mammography - ANN & Mammography - GROUNDTRUTH:GROUNDTRUTH:

All patients operated All patients operated and biopsy and biopsy reports reports availabale:availabale:- 54 malignant, 46 benign54 malignant, 46 benign

All patients rated to be All patients rated to be - benign, indeterminate, benign, indeterminate,

malignantmalignant Manual marking of Manual marking of

- 828 indiv. microcalcifications828 indiv. microcalcifications- 735 artifacts735 artifacts

ANN & Mammography ANN & Mammography - -

HARD & SOFT:HARD & SOFT:

ANN & Mammography ANN & Mammography - -

HARD & SOFT:HARD & SOFT: Hardware:Hardware:

- SunSparc20SunSparc20- Neurocomputer Synapse-1 (SNAT):Neurocomputer Synapse-1 (SNAT):

Software: Software: - Image Processing: IDL 5.0, Creaso Image Processing: IDL 5.0, Creaso

Research SystemResearch System- ANN - Synapse-1: ANN - Synapse-1: nneuronal euronal

AApplication pplication LLanguageanguage(SNAT, Germany) - C++ libraries (SNAT, Germany) - C++ libraries

ANN & Mammo - ANN & Mammo - WORKFLOWWORKFLOW

ANN & Mammo - ANN & Mammo - WORKFLOWWORKFLOW

Backgroundcorrection

Classification:

Feat.Extr. ANN

malign

benignindeterminate

ANNtypical

Detection:

ANNFeature Extraction

ANN & Mammography -ANN & Mammography -BACKGROUNDKORRECTIOBACKGROUNDKORRECTIO

NN

ANN & Mammography -ANN & Mammography -BACKGROUNDKORRECTIOBACKGROUNDKORRECTIO

NN

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION: Line feature:Line feature:

- Calculate Sobel GradientsCalculate Sobel Gradients- map to 16 directionsmap to 16 directions- make histogramm 9x9 make histogramm 9x9 - if difference between 2 peaks > 6/16 if difference between 2 peaks > 6/16

and < 10/16and < 10/16 multiply both peaks = line multiply both peaks = line

feature for a pointfeature for a point

Linefeatures:Linefeatures:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

Linefeatures:Linefeatures:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

Linefeatures:Linefeatures:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

DETECTION:DETECTION:

ANN & Mammography -ANN & Mammography -RESULTS DETECTION:RESULTS DETECTION:

ANN & Mammography -ANN & Mammography -RESULTS DETECTION:RESULTS DETECTION:

ANN & Mammography -ANN & Mammography -FEATURES FOR CLASSIFICATIONFEATURES FOR CLASSIFICATION

ANN & Mammography -ANN & Mammography -FEATURES FOR CLASSIFICATIONFEATURES FOR CLASSIFICATION

Mean vector of population: Mean vector of population: mmxx = = EE{{xx}}

Covariance matrix: Covariance matrix: CCxx = = EE{({(x x - - mmxx)()(x x - - mmxx))TT}}

Calculate eigenvectors of Calculate eigenvectors of CCx x , and , and transformationmatrix Atransformationmatrix A

Transform image: Transform image: yy = = AA((x x - - mmxx) - ) - Hotelling Hotelling TransformationTransformation

ANN & Mammography -ANN & Mammography -FEATURES FOR CLASSIFICATIONSFEATURES FOR CLASSIFICATIONS

ANN & Mammography -ANN & Mammography -FEATURES FOR CLASSIFICATIONSFEATURES FOR CLASSIFICATIONS

Extension in 8 directionsExtension in 8 directions „ „Minimum Enclosing Minimum Enclosing

Rectangle“:Rectangle“:- Center of gravity, excentricity, Center of gravity, excentricity,

aspect-ratio ....... aspect-ratio .......

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS

Features of individual Features of individual microcalcifications:microcalcifications:- Meanvalue of greylevels of pixels within Meanvalue of greylevels of pixels within

one mcone mc- Local contrast of one mcLocal contrast of one mc- Border gradientsBorder gradients- Area, perimeter and compactness (PArea, perimeter and compactness (P22/A)/A)

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS Clustering - recursive Clustering - recursive algorithmalgorithm

Area of the cluster - Convex Area of the cluster - Convex Hull ProcedureHull Procedure

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS Features from „Convex Features from „Convex

Hull“:Hull“:- Number of mc’sNumber of mc’s- Area, perimeter and compactness of Area, perimeter and compactness of

clustercluster- Density (number of mc’s/A)Density (number of mc’s/A)- Inter mc - distancesInter mc - distances- descriptive statistics of individual mc descriptive statistics of individual mc

featuresfeatures

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS

ANN & Mammography -ANN & Mammography -FEATURES FOR FEATURES FOR

CLASSIFICATIONSCLASSIFICATIONS Total features: n=73Total features: n=73 Automatic Automatic

selectionsprocess: selectionsprocess: patient based patient based „ Leave-one-Out-Test“ „ Leave-one-Out-Test“ for every featurefor every feature

10 suited for differentiation typical - 10 suited for differentiation typical - indeterminateindeterminate

10 suited for differentiation typical - 10 suited for differentiation typical - indeterminateindeterminate

12 suited for differentiation benign - 12 suited for differentiation benign - malignantmalignant

12 suited for differentiation benign - 12 suited for differentiation benign - malignantmalignant

Sensitivity : 97%Specifity : 34%

Sensitivity : 97%Specifity : 34%

Sensitivity : 98%Specifity : 47% Sensitivity : 98%Specifity : 47%