Local Feature Based Breast Tissue Appearance Modelling for Mammographic Risk Assessment

Many studies have indicated that there is a strong correlation between breast tissue density/patterns and the risk of developing breast cancer. Therefore, modelling breast tissue appearance in mammograms is important for automated mammographic risk assessment. In this paper, we present a method for building models of breast tissue appearance based on local features in mammographic images. Mammographic tissue is modelled based on statistical analysis of local tissue appearance. We investigate five strategies by employing different types of local features, describing aspects of intensity, texture, and geometry. A visual dictionary is generated to summarise local tissue appearance with descriptive “words”. The overall appearance of the breast is represented as an occurrence histogram over the dictionary. The resulting histogram models can be applied to breast density classification. The validity is qualitatively and quantitatively evaluated using the full MIAS database based on the BIRADS density classification. We test the performance of each individual strategy and the combination of all strategies. The best classification accuracy is 78% for the four BIRADS categories. This increases to 90% for two-class (low/high) density classification. The obtained results indicate that our method has potential for mammographic risk assessment.