Segmentation of anatomical branching structures based on texture features and graph cut

Segmentation of tree-like structure within medical imaging modalities, such as x-ray, MRI, ultrasound, etc., is an important step for analyzing branching patterns involved in many anatomic structures. However, images acquired using these different acquisition techniques frequently have features of poor contrast, blurring and noise, and therefore the segmentation result of traditional image segmentation methods may not be satisfactory. In this paper, we propose a framework for accurate segmentation of the ductal network in x-ray galactograms. Our approach is based on the graph cut algorithm and texture analysis to extract features of skewness, coarseness, contrast, energy and fractal dimension. The features are chosen to capture not only architectural variability of the enhanced ductal tree, but also spatial variations among pixels. The proposed approach was applied to a dataset of 20 galactographic images. We performed receiver operating characteristic (ROC) curve analysis to assess the accuracy. The area under the ROC curve observed was 0.76, indicating that our approach may potentially assist clinicians in the interpretation of breast images and facilitate the investigation of relationships among structure and texture of the branching patterns.