Image unwarping and difference analysis: a technique for detecting abnormalities in mammograms
Maha Sallam · 1997
Corresponding images are used by radiologists as a source of information for determining the existence of abnormalities in mammograms. The images may be time sequences of the same breast from two different screening exams, or they may be bilateral images of the left and right breasts obtained during the same session. A fully automated mammogram comparison technique is proposed for identifying differences between corresponding images. The technique recovers an approximate measure of the deformation between a pair of mammograms based on identifying corresponding features across the two images. Two types of features are used: object boundary features, and texture features. Matching of object boundary features is more tolerant to large deformations, and hence these features are used to recover most, of the global component of the deformation. The texture features are more difficult to match. However, their matching becomes more reliable by removing the global deformation recovered using boundary features. The texture features help recover more of the local component of deformation by virtue of having a better distribution throughout the image. Once an estimate of the deformation between an image pair is recovered, the registration process is completed using an unwarping technique for transforming one image into the coordinate system of the other. A difference image between the two registered images is generated using intensity-weighted subtraction in order to identify dense regions of large difference which often characterize potential abnormalities. Rigorous evaluation of the mammogram subtraction technique, and the extent to which the difference image reveals information about potential abnormalities, is performed using 145 bilateral image pairs which contain a total of 77 abnormalities of different types. A small set of 8 pairs of abnormal time sequences of mammograms is also used in the evaluation. The mammogram registration technique was successful in generating difference images in which at least 80% of the abnormalities were clearly identified by simple thresholding. The degree to which abnormalities are detectable by the proposed technique differs with the type, size and subtlely of the abnormality. Undetected abnormalities are analyzed in order to understand how corresponding image analysis can be complemented by other techniques in fully automated mammogram analysis systems.