CALMA (Computer Assisted Library for MAmmography), a collaboration among physicists and radiologists, has collected a large database of digitized mammographic images (about 5000) and developed a CAD (Computer Aided Detection) which has been integrated in a station which can be used also for digitization, as archive and to perform statistical analysis. In this work we present the results obtained in the automatic search of microcalcification clusters. Images (18×24 cm2, digitized by a CCD linear scanner with a 85 μm pitch and 4096 gray levels) are fully characterized: pathological ones have a consistent description with radiologist's diagnosis and histological data; non pathological ones correspond to patients with a follow up of at least three years. The automated microcalcification clusters analysis is made using an hybrid approach containing both algorithms and neural networks by which are extrated the ROIs (Region Of Interest). These ROIs are indicated on the images and a probability of containing a microcalcification cluster is associated to each ROI. The results obtained with this analysis are described in terms of the ROC (Receiver Operating Characteristic) curve, which shows the tree positive fraction (sensitivity) as a function of the false positive fraction (1-specificity) obtained varying the threshold level of the ROI selection procedure.
Search of microcalcification clusters with the CALMA CAD station
GOLOSIO, BRUNO;
2002-01-01
Abstract
CALMA (Computer Assisted Library for MAmmography), a collaboration among physicists and radiologists, has collected a large database of digitized mammographic images (about 5000) and developed a CAD (Computer Aided Detection) which has been integrated in a station which can be used also for digitization, as archive and to perform statistical analysis. In this work we present the results obtained in the automatic search of microcalcification clusters. Images (18×24 cm2, digitized by a CCD linear scanner with a 85 μm pitch and 4096 gray levels) are fully characterized: pathological ones have a consistent description with radiologist's diagnosis and histological data; non pathological ones correspond to patients with a follow up of at least three years. The automated microcalcification clusters analysis is made using an hybrid approach containing both algorithms and neural networks by which are extrated the ROIs (Region Of Interest). These ROIs are indicated on the images and a probability of containing a microcalcification cluster is associated to each ROI. The results obtained with this analysis are described in terms of the ROC (Receiver Operating Characteristic) curve, which shows the tree positive fraction (sensitivity) as a function of the false positive fraction (1-specificity) obtained varying the threshold level of the ROI selection procedure.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
