Study On Detection And Classification Methods For Breast Cancer Based On Multi-view Mammograms

Breast cancer is the most common cancerous tumor among women. China has become one of the countries where the breast cancer morbidity is with the fastest speed. Moreover, breast cancer is getting more frequent among young people in China. Currently, the incidence reason and the prevention of breast cancer are not completely known. Early detection is the key to increase recovery rate and decease mortality rate for breast cancer. Among different breast imaging techniques, mammography is one of the most widely used methods to screen women without obvious symptom. Breast cancer detection in China has the following problems. Firstly, the detection result is more on subjective analysis of radiologist and less on objective and quantative analysis. Secondly, the number of mammograms in the population screening is large and interpreting these images is a heavy workload for radiologists. Thirdly, compared with the breasts of Western women, the breasts of Chinese women tend to be denser. The overlapped glandular tissue may present the similar characteristics with tumor or cover tumors, which makes the detection of breast cancer more difficult. Using the technique and theory of image processing and pattern recognition, objective and quantative results can be obtained, which contributes to detecting the breast cancer at an early stage.The existing detection methods for the breast cancer were normally applied to mammograms of Western women, which have more fatty-breast compared with Chinese women. When these methods were transplanted into cancer detection in mammograms of Chinese women, there exist some problems. Based on Western mammograms and Chinese mammograms, some key issues in breast cancer detection and classification were deeply studied in this thesis. The main research contents dealt with pectoral muscle segmentation, mass detection using multi-views, mass retrieval and mass classification. The innovation contributions of this thesis were summarized as follows:(1) Pectoral muscle segmentation:For pectoral muscles with homogenous texture, one method based on anatomical characteristics (homogenous texture and high intensity deviation) was proposed. The pectoral muscle edge map was defined based on anatomical characteristics. Then the initial pectoral muscle edge was computed through the edge map, thus the assumption of straight lines could be avoided. Finally, the initial edge was modeled as an accelerated displacement curve and refined by the Kalman filter, which provided a novel view for modeling the pectoral muscle. Experiment results showed that the proposed method achieved better segmentation results compared with existing methods. For pectoral muscles with complex texture, another method based on spectral clustering and region merging was proposed. Firstly, spectral clustering combined with edge information was presented and used to segment the mammogram, leading to the segmentation of pectoral muscle more accurate. Then a region merging algorithm was proposed to merge the segmented regions according to the characteristics of triangle shape. The problem of segmenting the pectoral muscle with complex texture into multiple sub-regions could thus be solved. Experiment results showed that the proposed method was more effective in segmenting the pectoral muscle with complex texture.(2) Bilateral mass detection:One bilateral analysis method for mass detection was presented combining region matching by shape context with hierarchical similarity measurement. The matching cost was defined to quantatively evaluate matching credibility in the region matching method. The problem of discarding matching credibility was thus solved. As the existing measurements were not highly discriminating in mass and normal regions, a hierarchical similarity measurement considering both global feature and local feature was designed. Experiment results showed that the proposed mass detection method using bilateral analysis achieved better detection results in dense breasts, compared with detection using unilateral analysis and existing bilateral analysis methods.(3) Mass retrieval:One mass retrieval method was put forward combining discriminating anchor graph hashing (DAGH) with linear neighborhood propagation (LNP). On the basis of original AGH, pathology class information was introduced to compute the image similarity and DAGH was proposed as a new image representation. The pathology relevance was enhanced in DAGH representation. LNP was employed as the relevance feedback technique and the interactive retrieval for mammogram masses was implemented. Experiment results showed that without segmenting the mass contour the proposed method achieved better mass retrieval performance in dense breasts, compared with original AGH and other existing methods.(4) Mass classification:One method to classify mass as benign or malignant was presented based on the texton under unequal interval subsamples. Existing methods by texture descriptor artificially set the texture dictionary. To solve this problem, texton was selected as the texture feature, where the texture disctionary was obtained through the clustering analysis of the training images. Unequal interval subsample strategies were designed to capture different discriminating structures. Integrating all the discriminating structures, texton could be less scale dependent. Experiment results showed that the proposed method obtained better performances in mass classification without segmenting the mass contour, compared with the existing texture-based methods.