The aim of the experiments is to demonstrate the benefits of multiple expert fusion in image analysis. In particular the problem is to label, either as normal or as abnormal (Microcalcifications), all suspected regions which are detected in mammographic images by the segmentation method described in [ 1 ].
A set of 227 digitised mammograms including 22 images containing microcalcifications and 205 normal images was used in the experiment [ 5 ]. The images are part of the Mammographic Image Analysis Society (MIAS) database. The preprocessing step [ 1 ] outputs a set of 39 independent measurements for each segmented region. A training set comprising 320 single MCs from 3 abnormal images and 960 regions extracted from 5 normal images was then formed. A sequential floating search feature selection method [ 4 ] was applied to select four subsets of features (FS1, FS2, FS3 and FS4) out of the 39 available measurements. The four feature sets were constrained to be complementary as far as practicable i.e. to constitute distinct object representations. Nevertheless we found that four features were shared by all four classifiers. However in [ 3 ] we demonstrated that the sum rule combination still remains a justifiable strategy. By computing an average within class conditional matrix we confirmed that the correlations between all the other features were low and therefore the assumption of independence could be deemed to hold.
The 8 images used for training were then excluded and the remaining images were divided into two sets, Set A and Set B. Set A contained 10 abnormal images and 100 normal images and Set B is made up of 9 abnormal and 100 normal images.
S Ali Hojjatoleslami