Segmentation Of Masses In MRI Breast Imaging

Breast cancer is the most prevalent cancer with high mortality rate in middle-aged women ofworldwide. Currently, the pathogenesis of breast cancer has not been completely clarified. The mosteffective treatment strategy for breast cancer is still “early detection and early treatment”. Breastimaging is one of the major tools for early detection of breast cancer. Because the dynamic contrastenhanced magnetic resonance imaging (DCE-MRI) shows a higher sensitivity in the detection ofearly dense breasts cancer compared with other imaging modalities. DCE-MRI is widely used todiagnose and evaluate the treatment of breast cancer. Currently, diagnosis of breast MRI imagesmostly relies on personal experience and expertise of clinicians. In order to reduce the misdiagnosisrate and improve the efficiency of diagnosis, computer-aided diagnosis (CAD) of breast MRItechnology attracted the attention of researchers. Breast mass segmentation as an important part ofbreast MRI CAD system, its quality will affect the feature extraction and classification which arebased on the segmentation result. However, due to the complexity of the breast tissue, the limitedresolution of the breast MRI image, partial volume effect, bias field and other factors, theboundaries between lesion and surrounding tissue are often blurred in breast MRI image. Theperformance of exist MRI image segmentation algorithm of breast lesions has not reached a verysatisfactory level.This study proposes a parameter adaptive three dimensional segmentation method for breastDCE-MRI tumor from3D and parametric adaptive perspective based on the characteristics of breastDCE-MRI images. The segmentation of the whole process is consisted of two parts, coarsesegmentation and fine segmentation. Coarse segmentation is mainly to provide the label field forthe fine segmentation process. In fine segmentation process, in order to improve segmentationperformance, we combined the three-dimensional information and adopted parameter adaptivestrategy in this process. Combined with tree-dimensional information is to solve the segmentationproblem of lesion ends. And used of parameter adaptive strategy is to settle the problem of loweraccuracy when segment the both ends of the lesion. Since this method combines the3D informationand adopts parameter adaptive strategy in the segmentation process, it not only can increase theaccuracy of ends segmentation, but also keep that of middle part. Moreover, the entire segmentationprocess is an automated process, so, the segmentation results are objective.In order to evaluate the propose method, we combined overlap rate method and “unsupervised”method to reach this goal. The evaluation result shows that segmentation method proposed in this study has high segmentation accuracy. Meanwhile, we used the segmentation method in tumorclassification experiment to further evaluate our method. In the experiment we extracted thedynamic enhancement features, morphological characteristics, statistical and texture features, totalof17-dimensional feature, based on the segmentation result. Then, we used the method ofSVM-RFE to select feature and SVM classifier to classify tumors. Classification resultdemonstrates that the features extracted from segmented regions have a good classificationperformance, and supports the truth that the proposed method has good segmentation performance.