Detection of masses in digitised mammograms

A method to detect masses in digitised mammograms based on textural information is proposed in this paper. Enhancement and segmentation of the images, based on texture is performed using wavelets, steerable filters and Laws textures maps. Hexagonal wavelets are proposed in which the sampling system exhibits the tightest packing of all regular two-dimensional sampling grids. In hexagonal sampling, orientations are partitioned into three bands of 60 degrees, equally covering the frequency domain such that no particular orientation is favoured. The so called lifting scheme is used to construct the wavelets. This scheme has several advantages over the conventional Fourier approach. Most importantly it can easily be extended to uneven sampling, and is not susceptible to boundary conditions. The lifting scheme is also computationally more efficient by a factor of two. Analysis was also performed with Laws filter masks. These can be used to segment or classify an image using textural features. The filter masks extract secondary features from the natural microstructure characteristics of the image. Another method discussed are oriented filters. Here filters are applied with arbitrary orientation and phase, and the output examined. A filter can be efficiently steered along the dominant orientation of the image, enhancing the underlying microstructure. The mentioned schemes were used for segmentation and enhancement. It was found that Laws gave a high true positive (88%) and false positive rate. Other methods, such as wavelet enhancement significantly reduced the false positive rate, but sacrificed true positives (68%). In this application it is equally as important to reduce the false detection as to find all tumours, since false detections could distract the radiologist.