The CALMA project A CAD tool in breast radiography A.Ceccopieri, Padova 9-2-2000.

The CALMA project A CAD tool in breast radiography A.Ceccopieri, Padova Computer Assisted Library in MAmmography Screening mammography sensitivity (identified positives / true positives) 73% - 88% specificity (identified negatives / true negatives) 83% - 92% These merit figures INCREASE if diagnosis is performed by 2 independent radiologists CALMA aims to: Build a DATABASE of mammograms in digital format Perform an automatic classification of parenchyma structures Detect the spiculated lesions Detect micro-calcification clusters 900 patients 2900 images Glandular 58 % DN 5 % FA 37 % OUR DATABASE DAQ: granularity: 85 m range:12 bit dimensions: ~2000x2600 pixels STORAGE 60 images/ CD (no compression) up to 240 CD HARDWARE DAQ panel & database search Queries Full screen display Preview and images description Automatic classification of breast parenchyma Left to right / top to bottom: - dense (DN) - irregularly nodular (IN) - micro-nodular (MN) - fiber-adipose (FA) - fiber-glandular (FG) - parvi-nodular (PN) -Glandular (IN+MN+FG+PN) Supervised FF-ANN Spatial frequencies analysis (FFT) 512x512 pixels analysis ANN classification Feature extraction 2dim FFT GLANDULAR RESULTS: TEXTURE ANALYSIS DENSE ADIPOSE GLANDULAR DENSE >95% 0% 0% ADIPOSE 16% 683% 16% GLANDULAR 4% 3% 931% SPICULATED LESIONS Unroll spirals Spatial frequencies analysis(FFT) FF-ANN examples MethodArea (cm2) spread (cm2) B (0-0)31 16 B (1-3)27 13 B (2-5)25 13 C neural36 12 C normalized36 18 C corona49 27 sensitivity=90(3)%: Integration range 2-5 Spiculated lesions: CAD performances Red= radiologist Blue= CAD RESULTS: SPICULATED LESIONS Sensitivity (per patient) 903% FALSE POSITIVES / IMAGE 1.4 AVERAGE ROI 25 cm 2 DATA REDUCTION ~ 10 Examples MICROCALCIFICATION CLUSTERS FF-ANN + Sanger learning rule PCA Method Image Preprocessing (convolution filters) PCA through a NN trained with the Sanger rule Study of the first Principal Components Classification Preprocessing 60x60 pixels windows selection convolution filters with dims: 5x5 7x7 9x9 Best results with a 7x7 filter with A=1\N 2 a ij