Classification of microcalcifications in digitised mammograms using multiscale statistical texture analysis

We present a multiscale statistical approach to texture analysis. These techniques are used to classify microcalcifications in digitised mammograms as benign or malignant. In this study we extract the proposed multiscale statistical texture signatures, based on the co-occurrence matrix, as well as wavelet-based texture signatures from the regions of interest containing the microcalcifications. The discriminatory ability of these texture signatures is demonstrated by their ability to successfully distinguish between benign and malignant cases. Classification is performed by means of a k-nearest neighbour classifier. One hundred percent correct classification is achieved when using a combination of the multiscale statistical texture signatures and the wavelet-based texture signatures. A database with a small number of samples was used, and further analysis with a larger database will give these results greater statistical significance.