Breast Mass Segmentation In Mammograms Based On Graph Cuts Algorithm

Breast cancer remains a major cause of cancer deaths among women around theworld. Mammography is currently accepted as the most effective measure for earlydetection and diagnosis of breast cancer. In order to improve efficiency and accuracyin mammographic breast cancer detection/diagnosis, the computer-aided diagnosis(CAD) technology has been widely used by radiologists. As an important step inbreast X-ray image CAD system, segmentation result of the mass reflects pathologicfeatures of the mass and directly affects subsequent feature extraction andclassification. However, there are many noise and artifact in the mammogram, andthere is low contrast between complex mass and the background. Existing methods inmass segmentation are often insufficient to meet clinical requirements.Graph Cuts has been used in recent years for interactive image segmentation.The core ideology of Graph Cuts is to map an image onto a network graph, andconstruct an energy function on the labeling, and then do energy minimization withcombinatorial optimization techniques. This study proposes a new segmentationmethod using iterated Graph Cuts based on multi-scale smoothing. The multi-scalemethod can segment mammographic images with a stepwise process from global tolocal segmentation by iterating Graph Cuts. An unsupervised watershed transform ofthe morphological gradient of the original X-ray image is used in our work to producea region adjacency graph for the optimization steps. Moreover important contours ofthe mammographic image are preserved during the first unsupervised segmentation.The new segmentation strategy effectively improves segmentation performance withless influences of image noise level.To improve insufficiencies of the multi-scale segmentation method, such asglobal minimization of the energy functions is NP-hard even in the simplestdiscontinuity-preserving case; this thesis study continues to introduce a fast, automaticand efficient segmentation method. Fuzzy-C means clustering algorithm is used forpre-segmentation of the mass to provide an initial mark field for the algorithm, andthenα- expansion algorithm based on Graph Cuts quickly searches local minimumof the energy function, and conducts fast approximate energy optimization. The proposed method improves calculation efficiency greatly while keeping segmentationprecision. The experimental results demonstrate that the proposed method can provideaccurate and highly robust segmentation results for mammographic image withintensity inhomogeneity and heavy noise.