Research On Methods For Medical Image Fusion Using Sparse Representation And Segmentation Using Level Set

Modern medical imaging has made important contributions to the prevention and treatment of human diseases,which is a milestone in the history of natural science.The emerging medical imaging technologies provide new tools for the imaging description of biological tissues,which play an important role in clinical diagnosis,surgery,radiotherapy,curative effect evaluation and so on.However,limited by the imaging mechanism,there are still some deficiencies in the medical images obtained by various medical imaging equipment,such as low spatial resolution,lack of structural information,insufficient metabolic information,or the image needs multimodal fusion to meet the accurate assessment of complex diseases.The existence of these problems not only provides power for the development of new medical imaging equipment,but also provides a useful place for the improvement of medical image processing.Medical image processing plays an important role in improving the quality of medical images,completing the complementary fusion of multi-modal medical images,and realizing the segmentation of important parts.Based on theories and methods such as sparse representation(SR),level set and active contour model(ACM),new algorithms for multimodal medical image fusion and segmentation are studied in this dissertation.The main innovations are as follows:(1)Aiming at the interpretability problem of the activity level measurement feature parameter(the L1 norm of the patch sparse coefficient)in SR-based fusion,a salient feature map reconstruction algorithm based on the absolute value vector of the patch sparse coefficient is proposed,which graphically shows the saliency characteristics corresponding to the vector.Furthermore,referring to the concept of generalized sparse representation,the above-mentioned feature parameter is defined as patch singularity,and a physical explanation of it in the field of SR-based fusion is given.Based on the above,a singular feature map reconstruction algorithm based on the maximum absolute value of the patch sparse coefficient is proposed,which clarifies that the L∞ norm of the patch sparse coefficient can be used to locate local singularities of images.The research results show that the patch sparse coefficient can be effectively used in image fusion and edge detection.(2)Aiming at the low efficiency of the second-order iterative batch dictionary learning fusion algorithm,a medical image fusion algorithm based on online dictionary leaming(ODL)and pulse coupled neural network(PCNN)is proposed.The algorithm uses the ODL algorithm with excellent performance and high efficiency to train the dictionary,and determines the fusion coefficient based on the PCNN,and achieves high-efficiency fusion under the premise of ensuring the fusion quality.Aiming at the problem of multi-modal medical image fusion under noise,a medical image fusion algorithm based on joint sparse representation(JSR)is proposed.With the help of the denoising function of JSR,the algorithm can effectively achieve fusion even in the presence of noise,thereby solving the problem of insufficient anti-noise ability of traditional fusion algorithms.Aiming at the problem that a single dictionary cannot reflect the differences of various image patches,inspired by the good performance of the clustering SR in image inverse problem,a medical image fusion algorithm based on SR of classified patches is proposed.The algorithm uses the characteristics of cross-location and non-local self-similarity of classified image patches to train a dictionary with good non-local features to preserve the edge,texture,and details of the source image,so as to obtain good subjective and objective performance.(3)Aiming at the insufficient utilization of feature information in single-modal ultrasound(US)medical image segmentation,an ACM(named GLM)based on variational level set is proposed,which combines global and local maximum class separation distance criterion.Further,mining the local entropy features of US images,a two-stage accurate segmentation algorithm is proposed.In this algorithm,first,the robustness of local entropy against local grayscale disturbances is used to realize automatic pre-segmentation.Second,the pre-segmentation result is used as the initial contour of GLM for precise segmentation.Experiments show that by making full use of the effective feature information of single-modal images,high-quality segmentation can be achieved.In the problem of PET-CT lung tumor medical image segmentation,the doctor’s clinical experience is integrated into the algorithm design,referring to the experience in manual delineation of the lung tumor,a PET-CT lung tumor fusion image segmentation method based on level set is proposed.This method combines the region scable fitting(RSF)model and the maximum likelihood classification criterion to establish a hybrid active contour model,which realizes the accurate segmentation of PET-CT lung tumors,and provides effective computer-assisted segmentation results for clinical diagnosis and treatment.