Texture analysis and boundary refinement to outline mammography masses

In studying mammograms specific features are sought in routine examinations as common indicators of malignancy. These include masses of an approximately regular spherical or spiculated shape. Here, the authors have attempted to apply established methods of texture determination within a structured sequence of image processing steps and examine the benefits of this systematic approach by comparing the outlines of regions with characteristic texture as presented by the authors´ method with those determined by expert opinion. If successful, such a technique would be valuable in determining changes in lesion size and shape after conservative therapy in a robust automatic manner. It was found that, under the test conditions employed, the filter and associated image processing sequence was able to produce outlines of both regular masses and spiculated lesions which were well matched to those of an expert, though the technique was unable to distinguish between the two lesion types. The method has a low response in nonsuspicious instances. The protocols were evaluated on small image extracts in the vicinity of the lesion, and then extended to large extracts containing the whole breast and associated tissue regions. The results for the small extracts were good in terms of regions highlighted. For the subsequent application to large extracts the results were also promising but not as selective. This work indicates that pre-processing with the texture matrices and subsequent texture energy determination yields more accurate segmentation of objects than could be achieved by pure intensity thresholding of the original image. As part of this procedure, histogram equalisation has been shown to be valuable in modifying the data prior to the texture analysis process allowing improved segmentation as well as aiding the visualisation of the quality of the resultant images. A second thresholding step is important as a method of eliminating all the dark background areas. The use of an active contour model as a boundary refinement tool has been shown to improve the match for the regular masses, but less so for the more complex shaped spiculated lesions. Its disadvantage is that it introduces several new parameters, values of which appear to be lesion specific