Common database for research in mammographic image analysis

Department of Diagnostic Radiology, University Hospital Nij megenThe NetherlandsComparison of different methods described in the literature to automatically analyze digital mammograms isdifficult, because the test sets these methods have been applied to may have differed very much in degree of difficultyand quality. Therefore, development of a common database of representative mammographic images will be of greatbenefit for researchers working in the field. There are a number ofissues to be discussed before deciding on the designof such a database, like 1) spatial and contrast resolution, 2) film digitization versus digital acquisition, 3) imagequality and calibration, 4) annotation of verified malignant features, and 5) the choice of images. Some remarks onthese issues will be put forward below.Spatial resolution is considered to be a major problem in digital mammography. Digital systems for acquisitionand dsplay are roughly limited to 2k2, currently, which means a pixel size of about 0.1 mm for a mammogram,and a spatial frequency limit of 5 lp/mm. This is much less than the resolution of 15 lp/mm which is required formammographic film/screen systems. It should be realized, however, that the value of resolution limit determinedwith line pattern phantoms is questionable for mammography. In two ROC studies we have compared the diagnosticaccuracy of radiologists in detecting lesions [1] and characterizing microcalcification clusters [2] on film and digitized