Study On Detection And Classification Methods Of Breast Masses Based On Mammograms

Breast cancer is a major health concern for women. Its detection in early stage helps to introduce proper medical interventions to cure the disease. Multiple imaging methods have been used for early breast cancer detection, the mammography is the most reliable and convenient approach. However, breast cancer at the early stage in mammograms is often obscured and accurate diagnosis is subject to radiologists’ expertise and experience. Therefore, false positive and false negative results are commonly appeared. Computer-aided diagnosis system of breast cancer mammography can assist radiologists to improve the accuracy, efficiency and consistency of breast cancer diagnosis. Nevertheless, in some situations, the performance of existing methods could not reach the requirements of the radiologists. To further improve the performance of mass detection and the accuracy of breast cancer diagnosis in mammogram, in-depth studies are carried out in this dissertation. The main contributions and innovations are as follows:1. To improve the speed of mass detection, one mathematical modeling method of masses based on mammographic characteristics is proposed. Using this model, image features were defined to localize breast cancer masses quickly. On this basis, a clustering algorithm was employed to obtain the whole mass area.2. A novel mass detection method based on Marker Simplified Pulse Coupled Neural Network (Marker-SPCNN) is proposed aiming at the shortcomings of the existing mass layered detection algorithm. On the premise of guaranteeing a lower false positive rate, the detection sensitivity is improved effectively.3. For the purpose of segmenting mass precisely, a novel mass segmentation method based on SPCNN and modified Vector Chan-Vese (Vector-CV) model is proposed. Compared with existing methods, this approach is more suitable for processing Oriental female’s mammograms with low contrast.4. A new morphological characteristic based on bifurcation points in this skeleton is proposed, which makes the mass shape description more detailed. Moreover, a texture feature extraction method is put forward by combining Undecimated Wavelet Transform (UWT) with Gray Level Co-occurrence Matrix (GLCM). Experimental results show that the features extracted by proposed mehtod have a significant advantage in distinguishing benign mass and malignancy. 5. A Feature Weighted Support Vector Machine (FWSVM) based mass classification method is purposed to solve the problem that existing SVM based approach sets all the weights of features the same. On the basis of allocating feature weights reasonably, the FWSVM classifier is used to distinguish malignancy and benigncy. The experimental results show that the performance of this method is better than existing approaches.In this dissertation, Interdisciplinary studies across information engineering and biomedical science is carried out. The methods developed here helped improve some key steps in computer-aided diagnosis of breast cancer mammography. Promising qualitative and quantitative results obtained in this dissertation indicate that these methods can improve the accuracy and efficiency of breast cancer diagnosis, better meet the radiologists’ requirements in decision-making process, and may have significant impact on breast cancer early detection and women health at large.