The CALMA project A CAD tool in breast

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

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

FA 37 % OUR DATABASE DN 5% 900 patients 2900 images Glandular 58 %

HARDWARE DAQ: granularity: 85 mm range: 12 bit dimensions: ~2000 x 2600 pixels STORAGE 60 images/ CD (no compression) up to 240 CD

DAQ panel & database search Queries Preview and images’ description Full screen display

Automatic classification of breast parenchyma Spatial frequencies analysis (FFT) 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

2 dim FFT Feature extraction 512 x 512 pixels analysis ANN classification GLANDULAR

RESULTS: TEXTURE ANALYSIS DENSE ADIPOSE >95% 0% GLANDULAR 0% ADIPOSE 16% 68± 3% 16% GLANDULAR 4% 3% 93± 1%

SPICULATED LESIONS Unroll spirals Spatial frequencies analysis(FFT) FF-ANN examples

RESULTS @ sensitivity=90(± 3)%: Method Area (cm 2) spread (cm 2) B (0 -0) 31 16 B (1 -3) 27 13 B (2 -5) 25 13 C neural 36 12 C normalized 36 18 C corona 49 27

Integration range 2 -5

Red= radiologist Blue= CAD Spiculated lesions: CAD performances

RESULTS: SPICULATED LESIONS Sensitivity (per patient) 90± 3% FALSE POSITIVES / IMAGE 1. 4 AVERAGE ROI 25 cm 2 DATA REDUCTION ~ 10

MICROCALCIFICATION CLUSTERS FF-ANN + Sanger learning rule PCA Examples

Method • Image Preprocessing (convolution filters) • PCA through a NN trained with the Sanger rule • Study of the first Principal Components • Classification

Preprocessing • 60 x 60 pixels windows selection • convolution filters with dims: 5 x 5 7 x 7 9 x 9 Best results with a 7 x 7 filter with A=1N 2 aij <0 element) (aij kernel

Results No Microcalcification clusters With microcalcification clusters Sensitivity = 73 ± 2 % Specificity= 94 ± 2 %

Red= radiologist Micro-calcification clusters: CAD Blue= CAD 3 2 1

RESULTS: MICRO-CALCIFICATION CLUSTERS SENSITIVITY 73± 2% SPECIFICITY 94± 2%

FUTURE • Software developement: Local classification of parenchyma 2 Use parenchyma classification for lesions CAD 3 - Use the asymmetry between the two sides to detect cancer. • Increase the DATABASE • “ON-LINE Validation”: Is CALMA a good (second) radiologist? • Implementation of physicianfriendly CAD workstations in the collaborating Hospitals 1 -