Computer-aided Mass Diagnosis On Mammograms

Breast cancer is a malignant tumor disease with high incidence among women. Early detection and treatment are essential to reduce the mortality. Mammography is the dominant method for early detection of breast cancer. With the development of medicine and computer technology, developing computer-aided diagnosis (CAD) technology to help clinicians to detect suspect lesions, such as microcalcification and mass, and provide diagnostic suggestions, which can help decreasing the missed diagnosis rate and unnecessary biopsies, has become the research hotspots of the early diagnosis for breast cancer. But in clinical application, due to the complex and changeable mass image feature, and vulnerable to the surrounding tissue disturbance, the diagnosis result of mass is poorer than microcalcification and it is hard to meet the needs of clinical application. Based on the reasons above, the purpose of this thesis is to study more accurate and effective CAD algorithms of mass to obtain better diagnostic accuracy. The contents are listed as follows:1. Image enhancement. Single fuzzy enhancement approach was firstly used to adjust the gray-scale range for improving the contrast between tumor region and background. Then Gaussian mask was used to reduce the interference of surrounding tissues. Finally the image was converted to Radon domain to enhance the edge of the mass.2. Mass region segmentation and spiculated patterns detection. An automatic random walks algorithm was proposed for mass region segmentation. Firstly,2D maximum entropy threshold, region growing algorithm and morphological method were used to get a series of labels automatically. Then the evaluation method of average edge gradient was used to select the effective labels for random walks segmentation, and the region segmentation result was obtained. After that, the spiculated pattern was also detected. The Cartesian coordination was converted to polar coordination, and the line structure was enhanced by fuzzy entropy approach and line detection templates, thus the spiculated pattern was separated from the background. Finally, we used the gold standard of mass sample to evaluate the segmentation result. Compared with other segmentation algorithms, such as region growing and active contour, the outcome of the proposed algorithm achieves higher segmentation accuracy.3. Features extraction and optimization. Firstly, the features were extracted from the segmented regions of mass including intensity features, shape features, texture features and patients' basic information. Then, the genetic algorithm (GA) was used for the feature optimization.4. Mass classification. Machine learning method was selected for the mass classification and the classification results of masses before and after feature optimization was compared. The results show that the method of feature optimization can improve the classification accuracy and the proposed algorithms of the thesis achieves more accurate and effective diagnosis results.