Study Of Computer-aided Detection Methods Based On Mammographic Images

Mammography is considered as the most reliable and effective method for earlyscreening and diagnosis of the breast cancer. It plays important roles in early diagnosis andtreatment for improving the cure rates and reducing the mortality rates of breast cancer.However, inspecting of the mammograms by radiologists is not only inefficient but alsosubject to cause misdiagnosis and missed diagnosis of the breast cancer. For tackling sucha problem, this thesis devotes itself in developing techniques for automatic detection of thebreast cancer from the mammographic images and constructing a computer-aideddiagnosis system. The main work and contributions of this thesis are as follows:In order to suppress the noise and extract the mass correctly, a novel mammographicimage denoising method is proposed based on the non-subsampled contourlet transform(NSCT) and the symmetric normal inverse Gaussian (SNIG) model. In the framework ofBayesian maximum a posteriori estimation, the problem of denoising is reduced to aproblem of threshold derivation. A novel strategy is then proposed to determine thethreshold that is not only adaptive to different directions and scales, but also able to takeinto considerations the scale-to-scale difference in the contribution of the NSCTcoefficients to the noise. Experimental results show that the proposed method can not onlyreduce the noise but also well preserve edges and fine details in the mammographicimages.In order to reduce the effect of the background on the mass detection procedure, anovel adaptive thresholding method based on wavelet transform is proposed to segment thebreast region from the mammograms. First, a two dimensional wavelet transform isperformed with respect to the mammogram to alleviate the noise susceptibility caused bynonstationary distributions of intensities in the mammographic images. Then a onedimensional transform is performed with respect to the selected low frequency image for reducing the fluctuations, after which the local minima in the histogram curve is found anda threshold is then determined for extracting the breast region from the background.Experimental results indicate that the breast contours extracted by our method are well inconsistence with the corresponding ground truth.In order to reduce the redundant information in the mammographic images andimprove the speed and accuracy of the mass detection system, a ROI (region of interest)extraction method is proposed based on the adaptive threshold value interval and accordingto the hypothesis that mass growth can produce concentric layers. Using the concentriclayer rules and morphological features, suspicious mass regions areas are automaticallydetected and the ROIs are then determined. Experimental results show that our method canachieve higher sensitivity and lower false positive rate.In order to extract the contour of the mass from the extracted ROI, a novel masssegmentation method is proposed based on graph theoretic isoperimetric algorithm and thegradient vector flow (GVF) snake model. In this method, the isoperimetric algorithm isemployed to perform a rough segmentation of the mass and provide the initial position forthe GVF snake model, while the GVF snake model is used to obtain more accurate contourof the mass. Experimental results show that our method is more accurate and can beimplemented faster than other methods.