Medical image fusion has very important application value for medical image analysis and diseases diagnosis.According to the characteristics of multi modality medical image and human visual features,a new medical image fusion algorithm in NSCT (nonsubsampled coutourlet,NSCT) domain is proposed.Firstly,source images after registration are decomposed into low and high frequency sub-bands using NSCT.According to the low frequency subbands concentrating the majority energy of the source image and determining the image coutour,a fusion rule based on weighted region average energy combined with average gradient is adopted in low frequency subband coefficients.Moreover,according to human visual system which is more sensitive to contrast and edge,texture of image,the fusion strategy based on directive contrast integrated with the improved energy of Laplacian and PCNN (Pulse Coupled Neural Network,PCNN) are used to fuse high-frequency subbands.Furthermore,a closed loop feedback is introduced into the fusion rules of low and high frequency subbands to obtain optimal fused weights adaptively by using WSSIM (Weighted Structure Similarity,WSSIM) which highly consistent with the HVS(human visual features,HVS) as objective function.Finally,a lot of experiments of fusion of images including gray images and color images based on different fusion methods are conducted.The experiment results are analyzed in terms of visual quality and objective evaluation.The experiment results show that the proposed algorithm can effectively preserve information and significantly improve the performance of fusion image in terms of quantity of information,dispersed gray scale,visual quality and objective evaluation index.

Brain tumor segmentation from medical images is a clinical requirement for brain tumor diagnosis and radiotherapy planning.However,automatic or semi-automatic segmentation of the brain tumor is still a challenging task due to the high diversities and the ambiguous boundaries in the appearance of tumor tissue.To solve this problem,we propose a brain tumor segmentation method based on softmax regression and graph model.Firstly,the training samples are labeled from the multi-modality magnetic resonance images(MRI).Then,the softmax regression method is used to train the samples to obtain the parameters of this regression model and calculate the probabilities of each pixel belonging to different labels.At last,the probabilities calculated in the previous step are introduced to a graph-cut based model.This model is minimized with a min-cut/max-flow method to obtain the final tumor segmentation results.The experiment results demonstrate superior performance in brain tumor segmentation.

This paper presents an efficient and effective approach based on local shape structure classification for detecting media-adventitia border in intravascular ultrasound (IVUS) images.First,the category of local shape structures is found by using k-means clustering method.Second,patches from IVUS images indexed by the category are extracted by two kinds of features including integral channel and self-similarities features,and therefore a random decision forest model is constructed.Finally,the key points of testing IVUS images are detected using the trained classification model.Then with the help of curve fitting methods,detection of media-adventitia border is acquired.Experimental results demonstrate that the proposed algorithm effectively relieves the difficulties of interference factors such as plaques,artifacts and side vessel,and more accurately recognizes the key points of target border compared with existing algorithms,detects the whole target border successfully.The Jaccard Measure (JM) of media-adventitia border detected by the algorithm is 88.9%,Percentage of Area Difference (PAD) and Hausdorff Distance (HD) measures are reduced by 19.1% and 9.7% respectively.

With the rapid progress of precision medical technology, the segmentation of lesion regions in PET/CT images has played an important role in the development of medical plans. PET/CT combines two advanced imaging technologies organically: PET (functional metabolic imaging) and CT (anatomical structure imaging), which is an important progress in image diagnostics. Combined with the segmentation methods, this paper describes the characteristics of PET/CT images, the analysis of the current methods and the clinical application. Finally, the paper elaborates the development trend of PET/CT image segmentation technology.

Advances in medical imaging technologies and equipment play an important role in the biomedical researches.Cross-modality image-prediction technology predicts one modal image from that of another modal.This paper presents an overview of the literatures on medical imaging prediction technology and its applications,such as predicting Computed Tomography images from Magnetic Resonance (MR) images,7T-like MR image reconstruction,and predicting positron emission tomography images.The aim is twofold:the necessity and challenge for different modality medical image prediction technology;the overview and comparison of various methods in the field.We conclude that the cross-modality image prediction based on the deep learning technology has superiority in both predicting time and precision.

Guided image filtering is an effective edge-preserving smoothing technique. However, because the regularization parameter and amplification factor of the detail layer are fixed instead of content adaptive, the halo effects near the edges and noise in the background may be incurred, which will degrade the quality of the output images further. In this paper, an improved edge-aware weighting is first proposed to preserve the edges more accurately and reduce the halo effects. Then, the amplification factor of the detail layer is calculated in a gradient-directed way for boosting the details while suppressing the noise. Experimental results on the low-dose CT images with heavy noise and small details show that the proposed method can reduce the halo effects and attenuate the noise while enhancing the contrast efficiently. The visual quality of the output images meets the requirement of the clinical diagnosis.

To improve the accuracy of segmenting the tumor sub-regions in glioblastoma multiforme (GBM) multi-modal magnetic resonance (MR) images, a GBM multi-modal MR images automatic segmentation algorithm is proposed by using multi-class convolution neural network (CNN). Firstly, after 98% winsorization and registration for the GBM multi-modal MR images, the bias field was corrected by using the N4ITK method. Secondly, a multi-class CNN model mainly consisting of four convolutional layers, two pooling layers and two fully connected layers was constructed; the GBM multi-modal MR images were pre-segmented after training,and voxels were classified into five different labels.Finally,all false positive regions smaller than 200 voxels were removed,and the final segmentation results were obtained by median filtering. The Dice similarity coefficient DSC,positive predictive value PPV and average Hausdorff distance AHD were adopted as the evaluation index, and the DSC, PPV as well as AHD were 0.889±0.087, 0.859±0.127 and 1.923 for segmenting the entire tumor tissues in F-C-GBM dataset by the proposed algorithm, respectively. Results indicate that the proposed method can effectively improve the performance in the segmentation of the GBM multi-modal MR images and may be expected to have clinical application prospects.

Breast MR image segmentation is difficult because of complex organization and intensity inhomogeneity. This paper proposes a segmentation method based on dual-tree complex wavelet transform and density clustering. Firstly, the image is denoised by using complex wavelet domain bivariate model combined with anisotropic diffusion function; Then simple linear iterative clustering (SLIC) algorithm is used to obtain the neighbors of each superpixel, thereby reducing the time of searching for the nearest neighbor of each sample in KNN-DPC algorithm. Finally, nearest neighbor sample density information of superpixel region is introduced,and distribution strategies from KNN-DPC algorithm are used for adaptive clustering. The segmentation results of simulation and clinical data show that the proposed algorithm can segment breast MR images effectively.

Residual neural network (ResNet) has witnessed tremendous amount of attention in deep learning research over the last few years and has made great achievements in computer vision. In this paper, the ResNet is summarized in the following aspects: Firstly, the basic structure and working principle of the ResNet are expounded; Secondly, in model development, the eight network models of the ResNet are summarized in time sequence; Thirdly, in structural optimization, the research progress is described from five aspects of ResNet, including convolutional layer, pooling layer, residual unit, fully connected layer and the whole network; Finally, the application of ResNet in medical images processing is mainly discussed from two aspects of image recognition and image segmentation. In this paper, the principles, models, and structures of ResNet are systematically summarized, which has positive significance to the research and development of ResNet.

In this paper, we propose an improved Wasserstein generative adversarial network (WGAN), de-aliasing Wasserstein generative adversarial network with Gradient Penalty (DAWGAN-GP), for magnetic resonance imaging (MRI) reconstruction. This method uses WGAN to replace the traditional GAN, and combined the gradient penalty to improve the training speed and to solve the slow convergence problem of WGAN. In addition, for better preservation of the fine structures in the reconstructed images, we incorporate perceptual loss, pixel loss and frequency loss into the loss function for training the network. Compared with other state-of-the-art deep learning methods for MR images reconstruction, DAWGAN-GP method outperforms all other methods and can provide superior reconstruction with improved peak signal to noise ratio (PSNR) and better structural similarity index measure (SSIM).