Breast Cancer - FACTS: Mammography - TECHNIQUE: REPORTING: DILEMMA: Breast carcinoma leading cause of cancer death in womean Every 8-10

Transcription

1 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 in Austria Clustered mircrocalcifications one of early sign s Breast Cancer - FACTS: SURVIVAL vs TUMORSPREAD Total local regional Metastasis unstaged TECHNIQUE: REPORTING: Clsutered Microcalcifications are early sign s ofbreast cancer (bright spot s) - Size: : mm - search with a magnifying glass - Difficult perception in low contrast areas - Differentiate: cancerous vs non cancerous DILEMMA: Special training of radiologists necessary Positive predictive value of radiologists: : 20% Double Reading (independet( reporting by 2 radiologists): improves accuracy by 5-15% 1

2 Artificial Intelligence in HYPOTHESIS: Clustered microcalcifications be found reliably found by a computer application and a discrimination between cancerous and non cancerous microcalcifications be done using neuronal nets VISION CAD System (Computer( Aided Diagnosis) can act as never tired, second reader Variance of female FEATURES ANN & PATIENTS: Patient Database: patients = 272 Images with Mc s - High resolution film digitalization: 8000x6000x15 -> about 90 Mbyte/image - Resolution for image processing: : 91.5 µm ANN & GROUNDTRUTH: All patients operated and biopsy reports availabale: - 54 malignant,, 46 benign All patients rated to be - benign, indeterminate, malignant Manual marking of indiv. microcalcifications artifacts 2

3 ANN & HARD & SOFT: Hardware: - SunSparc20 - Neurocomputer Synapse-1 1 (SNAT): Software: - Image Processing: : IDL 5.0, Creaso Research System - ANN - Synapse-1: neuronal Application Language (SNAT, Germany) - C++ libraries ANN & Mammo - WORKFLOW Feat.Extr. Backgroundcorrection Detection: Feature Extraction Classification: ANN ANN typical ANN indeterminate malign benign ANN & BACKGROUNDKORRECTION ANN & Line feature: - Calculate Sobel Gradients - map to 16 directions - make histogramm 9x9 - if difference between 2 peaks > 6/16 and < 10/16 multiply both peaks = line feature for a point ANN & ANN & R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 R11 R12 R13 R14 R15 R16 Direction 3

4 ANN & ANN & RESULTS DETECTION: FEAT URE Az graylevel min max mean var edge gradients min max mean var line feature min max mean var local contrast (object) ,00 0,90 0,80 0,70 0,60 0,50 0,40 Az=0.96 Sensitivity = 0.90 Specificity = ,00 0,20 0,40 0,60 0,80 1, ANN & FEATURES FOR CLASSIFICATION Mean vector of population: m x = E{x} Covariance matrix: C x = E{( {(x - m x )(x - m x ) T } Calculate eigenvectors of C x, and transformationmatrix A Transform image: y = A(x - m x ) - Hotelling Transformation ANN & Extension in 8 directions Minimum Enclosing Rectangle : - Center of gravity, excentricity, aspect-ratio... ANN & Features of individual microcalcifications: - Meanvalue of greylevels of pixels within one mc - Local contrast of one mc - Border gradients - Area, perimeter and compactness (P 2 /A) 4

5 ANN & Clustering - recursive algorithm Area of the cluster - Convex Hull Procedure ANN & Features from Convex Hull : - Number of mc s - Area, perimeter and compactness of cluster - Density (number of mc s/a) - Inter mc - distances - descriptive statistics of individual mc features ANN & Total features: : n=73 Automatic selectionsprocess: patient based Leave-one one-out-test for 10 every feature suited for for differentiation typical - indeterminate Sensitivity :: 97% Specifity :: 34% 12 suited for differentiation benign - malignant Sensitivity :: 98% Specifity :: 47% 5