Lesions Segmentation And Features Extraction In Breast Dynamic Contrast-Enhanced MRI

Breast cancer is one of the leading causes of death in women. Early detection and treatment of breast cancer can reduce the mortality and morbidity. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has been an important tool because it can provide plentiful four dimension information to diagnose breast diseases. And computer-aided diagnosis (CAD) system based on DCE-MRI can assist radiologists for their quick and effective diagnosis and gradually be applied to clinical practice. Lesion segmentation and features extraction in breast DCE-MRI both are important contents to study CAD technologies. In this paper, a new active contour model only based on background information to segment three dimension (3D) tumors is proposed according to the difference between the intensity distribution of normal tissues and lesions. Then, many different features including kinetic, morphological, volume texture and spatiotemporal four aspects are extracted after segmentation. Finally, classification and feature selection are performed based on SVM. The main content of this paper is shown in the following paragraphs.First, accurate3D lesion segmentation is challenging since the lesions have complicated topological structures and different intensity distributions. Based on the situation referred above, the paper focuses on the lesion segmentation study. Combined with pathophysiological basis and clinical data, the paper found distributions of normal tissues around lesions are homogeneous, while distributions of lesions are different. This is an important basis for discrimination of benign and malignant lesions. The proposed algorithm is different from the methods unified modeling normal tissues and lesions. The paper first proposes an active contour model based on background-distribution to segment3D lesions.102lesions which cover all common types of breast tumors are segmented in this paper. Thus the generality of the proposed segmentation method is better. To compare the algorithm results, an experienced radiologist delineated all lesions to obtain artificial segmentation results as the ground truth. The results of the improved model and results generated by other different methods based on active contour model are compared with manual segmentation results. Finally, the performance of the proposed method is evaluated by several different evaluation parameters. Experimental results show that the improved model can solve the lesion segmentation weak boundary leakage and have a good robustness. In addition, the algorithm is simple and faster.Then, breast DCE-MRI can provide rich lesion imaging features, so the diagnosis needs to consider comprehensively different information from different aspects. Based on the extended two-dimensional (2D) features, the paper extracts fifty-six3D features including kinetic, morphological, volume texture, and spatiotemporal features from different angles and dimensions as identifying benign and malignant lesions foundation.At last, for the classification of benign and malignant lesions and the comparison problem between2D and3D features, the paper uses two class feature selection strategy. It uses the mean square distance as features sort criteria, and sequential forward selection to search features. And SVM is applied to class lesions and validate results. Three experiments including2D feature selection,3D feature selection and comparison between2D and3D features are performed. The experimental results show that the classification performance of3D features is better than one of2D features on the basis of all proposed features in this paper. And the combination of all different classes features need to be comprehensively investigated while studying on the lesions classification in breast DCE-MRI.