Automatic Detection Of Ultrasound Images Of Breast Tumors And Benign And Malignant Discrimination

Ultrasonography is one of the most frequently used methods for the early detection of breast tumors. It plays an important role in the clinical imaging diagnosis due to its efficiency, non-invasive, convenience and low-cost. A high performance computer-aided diagnosis (CAD) system is of great significance to further improve the accuracy of breast tumor detection and provide physicians with a dependable second opinion. However, due to the severe speckle noise, the poor image contrast, shadowing artifacts and irregular lesion shapes, it is challenging to build a fully automatic detection and classification system for breast ultrasonic images. In this dissertation, a novel and effective computer-aided method, including the region of interest (ROI) generation, breast tumor boundary extraction and classification, is proposed without any manual interference.In the first part, the ROI is generated automatically by the texture classification of pixels. Based on the denoised ultrasonic image, the local texture, the local gray level co-occurrence matrix and position features are firstly incorporated to measure each non-overlapping image grid. Then by taking the selected feature vector as the input of a self-organizing map (SOM) neural network, the candidate ROI and the background can be differentiated. Lastly, the final ROI is obtained by removing the influence of fake ROIs, the lesions which have similar characteristics. Thus, there is no need to mark out the ROI manually by physicians for the subsequent boundaries extraction.In the second part, a modified Normalized Cut (Ncut) method is proposed with the weighted gray values of neighborhood pixels, which succeeds in the automatic segment of the breast tumor in ultrasonic images. Firstly, the proposed method partitions the original ultrasound image into clusters with Ncut, and further obtains the initial contour of the tumor by employing the different gray values and spatial distributions of each cluster. Then, for a small proportion of inaccurate segmentation (e.g. grayscale leakage), a modified active contour model is used to adjust the initial boundary to obtain the final result. Compared with the traditional edge detection methods, the proposed method can accomplish the boundary extraction of tumors efficiently and automatically by involving less computation and requiring no manual interference (e.g. initializing contour manually). Thus, it is possible to further increase the automation degree of the computer-aided diagnosis.In the third part, three texture and five morphologic system-independent, robust and effective features are defined for each breast tumor; and a highly efficient Affinity Propagation (AP) clustering is used to fulfill the malignancy and benign classification for an existing database without any training process. Besides, the generalization capabilities of the AP clustering together with the traditional neural network tools including the back propagation artificial neural network (BPANN), the Fisher Linear Discriminant (FLD), and the support vector machine (SVM) are also analyzed to evaluate their classification performances.The proposed computerized diagnosis system is validated by 132 cases (67 benignancies and 65 malignancies) with its performance compared to traditional methods such as the level set segmentation, the artificial neural network classifiers, and so forth. Experiment results show that the proposed system, which is low in the computation complexity, training-free and highly automated, can accomplish the auto-detection and classification of ultrasonic breast tumors with the great accuracy. Thus, it is expected to be able to provide valuable references for the clinical diagnosis in the coming future.