Multiresolution Analysis And Segmentation Of Medical Images

Image segmentation is the technology of separating the foreground from the background of an image,detecting the target area from the entire area,enabling the next image processing operations such as image registration,fusion,image analysis and understanding,image semantic recognition,image 3D reconstruction,etc.Medical image segmentation is the process of segment a medical image into disjoint regions,such as organs,tissues,or lesions based on its characteristics.However,the computer-based medical image segmentation technologies have important research significance and high application values due to the complexity of medical images,low contrast,their vulnerability to noise,unclear boundaries,and the disability of the human eye to recognize certain features.In this thesis,we propose two segmentation methods based on multi-resolution analysis to solve the problem that the segmentation of medical images is easily affected by noise and inaccurate segmentation.First,a medical image is transformed in multi-resolution ways and decomposed into a series of images with different resolution.Using the image segmentation method,the low-resolution image is initially segmented,and a series of results obtained by initial segmentation are used as reference values to be applied to high-resolution images,and then an improved segmentation method is used to segment high-resolution images to achieve the purpose of filtering noise and precise segmentation.The main work of the thesis includes:?1?Improved Watershed algorithm segmentation based on wavelet transform.First,a medical image is decomposed using first-order wavelet transform into images with low-frequency approximations and high-frequency details.The threshold denoising is performed on the high-frequency details,and then the wavelet reconstruction is made using the low-frequency approximation components and the denoised high-frequency detail components,thus,an image with the same resolution of original image is obtained.Finally,the watershed segmentation with marker constraint is applied to the reconstructed image to obtain the segmentation result.?2?Fuzzy C-means clustering algorithm?FCM?segmentation based on image Gaussian Pyramid.First,Gaussian smoothing is performed on the medical image;then,the smoothed image is down-sampled to obtain an image pyramid in which the resolution from the bottom layer to the top layer is sequentially reduced.The top-level image is subjected to FCM segmentation,and the initial cluster center of the second-level image segmentation is initialized by using the cluster center of the top-level image segmentation.The cluster center of the lower image segmentation is initialized by the cluster center of the upper layer image in sequence until we obtain the segmentation result of the original image.Through subjective observation and quantitative evaluation of the experimental results,it is shown that the first method is better than the traditional marker-constrained watershed segmentation method.The number of regions is smaller,the contour is more precise,and the intra-region consistency index(5₀?7? is lower.The difference indicator DIR has a higher value.The first method proposed in this thesis effectively combines the advantages of multi-resolution medical images,retains most of the details of the original medical images,filters out most of the noise,avoids over-segmentation of images,and improves segmentation accuracy.Compared with the traditional FCM segmentation algorithm,the second method has fewer segmentation iterations and a shorter segmentation time.The second method proposed in this thesis extracts the main tissues more completely,shortens the image segmentation processing time,and improves the segmentation efficiency.In summary,the two segmentation algorithms proposed in this thesis produce improved results compared to the traditional segmentation algorithms,and they are more practical and effective.They can higher highlight the target area of medical images and have better application values.