Concerning the the size and morphological diversity of thyroid tissue and the complexity of surrounding tissue in thyroid ultrasound images, an ultrasound thyroid segmentation network based on feature fusion and dynamic multi-scale dilated convolution was proposed. Firstly, the dilated convolutions with different dilation rates and dynamic filters were used to fuse the global semantic information of different receptive domains and the semantic information in the context details with different ranges, so as to improve the adaptability and accuracy of the network to multi-scale targets. Then, the hybrid upsampling method was used to enhance the spatial information of high-dimensional semantic features and the context information of low-dimensional spatial features during feature dimensionality reduction. Finally, the spatial attention mechanism was introduced to optimize the low-dimensional features of the image, and the method of fusing high- and low-dimensional features was applied to retain the useful features of high- and low-dimensional feature information with the elimination of the redundant information and improve the network's ability to distinguish the background and foreground of the image. Experimental results show that the proposed method has an accuracy rate of 0.963±0.026, a recall rate of 0.84±0.03 and a dice coefficient of 0.79±0.03 in the public dataset of thyroid ultrasound images. It can be seen that the proposed method can solve the problems of large difference of tissue morphology and complex surrounding tissues.

In order to effectively solve the problems of not obvious boundaries, unclear abdominal organ display caused by high-frequency detail loss as well as the inconvenient application of single-model single-scale reconstruction in the super-resolution reconstruction of abdominal Magnetic Resonance Imaging (MRI) images, a multi-scale super-resolution algorithm based on parallel channel-spatial attention mechanism was proposed. Firstly, parallel channel-spatial attention residual blocks were built. The correlation between the key area and high-frequency information was obtained by the spatial attention module, and the channel attention module was used to study the weights of the channels of the image to the key information response degree. At the same time, the feature extraction layer of the network was widened to increase the feature information flowing into the attention module. In addition, the weight normalized layer was added to ensure the training efficiency of the network. Finally, a multi-scale up-sampling layer was applied at the end of the network to increase the flexibility and applicability of the network. Experimental results show that, compared with the image super-resolution using very deep Residual Channel Attention Network (RCAN), the proposed algorithm has the Peak Signal-to-Noise Ratio (PSNR) averagely increased by 0.68 dB at the×2,×3 and×4 scales. The proposed algorithm effectively improves the reconstructed image quality.

In order to improve the quality of medical image super-resolution reconstruction, a wide residual super-resolution neural network algorithm based on depthwise separable convolution was proposed. Firstly, the depthwise separable convolution was used to improve the residual block of the network, widen the channel of the convolution layer in the residual block, and pass more feature information into the activation function, making the shallow low-level image features in the network easier transmitted to the upper level, so that the quality of medical image super-resolution reconstruction was enhanced. Then, the network was trained by group normalization, the channel dimension of the convolutional layer was divided into groups, and the normalized mean and variance were calculated in each group, which made the network training process converge faster, and solved the difficulty of network training because the depthwise separable convolution widens the number of channels. Meanwhile, the network showed better performance. The experimental results show that compared with the traditional nearest neighbor interpolation, bicubic interpolation super-resolution algorithm and the super-resolution algorithm based on sparse expression, the medical image reconstructed by the proposed algorithm has richer texture detail and more realistic visual effects. Compared with the super-resolution algorithm based on convolutional neural network, the super-resolution neural network algorithm based on wide residual and the generative adversarial-network super-resolution algorithm, the proposed algorithm has a significant improvement in PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural SIMilarity index).

When noise and intensity distortion exist in brain images, the method based on structural information cannot accurately extract image intensity information, edge and texture features at the same time. In addition, the computational complexity of continuous optimization is relatively high. To solve these problems, according to the structural information of the image, a non-rigid multi-modal brain image registration method based on Improved Zernike Moment based Local Descriptor (IZMLD) and Graph Cuts (GC) discrete optimization was proposed. Firstly, the image registration problem was regarded as the discrete label problem of Markov Random Field (MRF), and the energy function was constructed. The two energy terms were composed of the pixel similarity and smoothness of the displacement vector field. Secondly, a smoothness constraint based on the first derivative of the deformation vector field was used to penalize displacement labels with sharp changes between adjacent pixels. The similarity metric based on IZMLD was used as a data item to represent pixel similarity. Thirdly, the Zernike moments of the image patches were used to calculate the self-similarity of the reference image and the floating image in the local neighborhood and construct an effective local descriptor. The Sum of Absolute Difference (SAD) between the descriptors was taken as the similarity metric. Finally, the whole energy function was discretized and its minimum value was obtained by using an extended optimization algorithm of GC. The experimental results show that compared with the registration method based on the Sum of Squared Differences on Entropy images (ESSD), the Modality Independent Neighborhood Descriptor (MIND) and the Stochastic Second-Order Entropy Image (SSOEI), the mean of the target registration error of the proposed method was decreased by 18.78%, 10.26% and 8.89% respectively; and the registration time of the proposed method was shortened by about 20 s compared to the continuous optimization algorithm. The proposed method achieves efficient and accurate registration for images with noise and intensity distortion.

In the traditional medical image fusion, it is necessary to manually set the fusion rules and parameters by using prior knowledge, which leads to the uncertainty of fusion effect and the lack of detail expression. In order to solve the problems, a Computed Tomography (CT)/Magnetic Resonance (MR) image fusion algorithm based on improved Generative Adversarial Network (GAN) was proposed. Firstly, the network structures of generator and discriminator were improved. In the design of generator network, residual block and fast connection were used to deepen the network structure, so as to better capture the deep image information. Then, the down-sampling layer of the traditional network was removed to reduce the information loss during image transmission, and the batch normalization was changed to the layer normalization to better retain the source image information, and the depth of the discriminator network was increased to improve the network performance. Finally, the CT image and the MR image were connected and input into the generator network to obtain the fused image, and the network parameters were continuously optimized through the loss function, and the model most suitable for medical image fusion was trained to generate the high-quality image. The experimental results show that, the proposed algorithm is superior to Discrete Wavelet Transformation (DWT) algorithm, NonSubsampled Contourlet Transform (NSCT) algorithm, Sparse Representation (SR) algorithm and Sparse Representation of classified image Patches (PSR) algorithm on Mutual Information (MI), Information Entropy (IE) and Structural SIMilarity (SSIM). The final fused image has rich texture and details. At the same time, the influence of human factors on the stability of the fusion effect is avoided.

Aiming at the ambiguity caused by the loss of details in the super-resolution reconstruction of medical images, a medical image super-resolution algorithm based on deep residual Generative Adversarial Network (GAN) was proposed. Firstly, a generative network and a discriminative network were designed in the method. High resolution images were generated by the generative network and the authenticities of the images were identified by the discriminative network. Secondly, a resize-convolution was used to eliminate checkerboard artifacts in the upsampling layer of the designed generative network and the batch-normalization layer of the standard residual block was removed to optimize the network. Also, the number of feature maps was further increased in the discriminative network and the network was deepened to improve the network performance. Finally, the network was continuously optimized according to the generative loss and the discriminative loss to guide the generation of high-quality images. The experimental results show that compared with bilinear interpolation, nearest-neighbor interpolation, bicubic interpolation, deeply-recursive convolutional network for image super-resolution and Super-Resolution using a Generative Adversarial Network (SRGAN), the improved algorithm can reconstruct the images with richer texture and more realistic vision. Compared with SRGAN, the proposed algorithm has an increase of 0.21 dB and 0.32% in Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM). It provides a deep residual generative adversarial network method for the theoretical research of medical image super-resolution, which is reliable and effective in practical applications.

Currently, the adaptivity of global training dictionary is not strong for brain medical images, and the "max-L ₁" rule may cause gray inconsistency in the fused image, which cannot get satisfactory image fusion results. A multi-modal brain image fusion method based on adaptive joint dictionary learning was proposed to solve this problem. Firstly, an adaptive joint dictionary was obtained by combining sub-dictionaries which were adaptively learned from registered source images using improved K-means-based Singular Value Decomposition ( K-SVD) algorithm. The sparse representation coefficients were computed by the Coefficient Reuse Orthogonal Matching Pursuit (CoefROMP) algorithm by using the adaptive joint dictionary. Furthermore, the activity level measurement of source image patches was regarded as the "multi-norm" of the sparse representation coefficients, and an unbiased rule combining "adaptive weighed average" and "choose-max" was proposed, to chose fusion rule according to the similarity of "multi-norm" of the sparse representation coefficients. Then, the sparse representation coefficients were fused by the rule of "adaptive weighed average" when the similarity of "multi-norm" was greater than the threshold, otherwise the rule of "choose-max" was used. Finally, the fusion image was reconstructed according to the fusion coefficient and the adaptive joint dictionary. The experimental results show that, compared with the other three methods based on multi-scale transform and five methods based on sparse representation, the fusion images of the proposed method have more image detail information, better image contrast and sharpness, and clearer edge of lesion, the mean values of the objective parameters such as standard deviation, spatial frequency, mutual information, the gradient based index, the universal image quality based index and the mean structural similarity index under three groups of experimental conditions are 71.0783, 21.9708, 3.6790, 0.6603, 0.7352 and 0.7339 respectively. The proposed method can be used for clinical diagnosis and assistant treatment.

Sparse coding similarity measure has good robustness to gray-scale offset field in non-rigid medical image registration, but it is only suitable for single-modal medical image registration. A non-rigid multi-modal medical image registration method based on multi-channel sparse coding was proposed to solve this problem. In this method, the multi-modal registration was regarded as a multi-channel registration, with each modal running in a separate channel. At first, the two registered images were synthesized and regularized separately, and then they were divided into channels and image blocks. The K-means-based Singular Value Decomposition ( K-SVD) algorithm was used to train the image blocks in each channel to get the analytical dictionary and sparse coefficients, and each channel was weightedy summated. The multilayer P-spline free transform model was used to simulate the non-rigid geometric deformation, and the gradient descent method was used to optimize the objective function. The experimental results show that compared with multi-modal similarity measure such as local mutual information, Multi-Channel Local Variance and Residual Complexity (MCLVRC), Multi-Channel Sparse-Induced Similarity Measure (MCSISM) and Multi-Channel Rank Induced Similarity Measure (MCRISM), the root mean square error of the proposed method is decreased by 30.86%, 22.24%, 26.84% and 16.49% respectively. The proposed method can not only effectively overcome the influence of gray-scale offset field on registration in multi-modal medical image registration, but also improve the accuracy and robustness of registration.

The dictionary training process is time-consuming, and it is difficult to obtain accurate sparse representation by using a single dictionary to express brain medical images currently, which leads to the inefficiency of image fusion. In order to solve the problems, a Computed Tomography (CT)/Magnetic Resonance (MR) brain image fusion method via improved coupled dictionary learning was proposed. Firstly, the CT and MR images were regarded as the training set, and the coupled CT and MR dictionary were obtained through joint dictionary training based on improved K-means-based Singular Value Decomposition (K-SVD) algorithm respectively. The atoms in CT and MR dictionary were regarded as the features of training images, and the feature indicators of the dictionary atoms were calculated by the information entropy. Then, the atoms with the smaller difference feature indicators were regarded as the common features, the rest of the atoms were considered as the innovative features. A fusion dictionary was obtained by using the rule of "mean" and "choose-max" to fuse the common features and innovative features of the CT and MR dictionary separately. Further more, the registered source images were compiled into column vectors and subtracted the mean value. The accurate sparse representation coefficients were computed by the Coefficient Reuse Orthogonal Matching Pursuit (CoefROMP) algorithm under the effect of the fusion dictionary, the sparse representation coefficients and mean vector were fused by the rule of "2-norm max" and "weighted average" separately. Finally, the fusion image was obtained via reconstruction. The experimental results show that, compared with three methods based on multi-scale transform and three methods based on sparse representation, the image visual quality fused by the proposed method outperforms on the brightness, sharpness and contrast, the mean value of the objective parameters such as mutual information, the gradient based, the phase congruency based and the universal image quality indexes under three groups of experimental conditions are 4.1133, 0.7131, 0.4636 and 0.7625 respectively, the average time in the dictionary learning phase under 10 experimental conditions is 5.96 min. The proposed method can be used for clinical diagnosis and assistant treatment.

The traditional fusion methods based on Pulse Coupled Neural Network (PCNN) have the shortcomings of too many parameters, the parameters and number of network iterations difficult to accurately set, and poor fusion effect. In order to solve the problems, a new image fusion algorithm using the connection item (L item) of Linking Synaptic Computing Network (LSCN) model was proposed. Firstly, the two images to be fused was input into the LSCN model respectively. Secondly, the L term was used to replace the ignition frequency in the traditional PCNN as the output. Then, the iteration was terminated by the multi-pass operation. Finally, the pixels of the fused image were obtained by comparing the values of L terms. The theoretical analysis and experimental results show that, compared with the image fusion algorithms using the improved PCNN model and the new model proposed on the basis of PCNN model, the fusion images generated by the proposed algorithm have better visual effects. In addition, compared with the fusion algorithm of LSCN using ignition frequence as the output, the proposed algorithm is all superior in edge information evaluation factor, information entropy, standard deviation, space frequency, average grads. The proposed algorithm is simple and convenient, which not only reduces the number of parameters to be determined, reduces the computational complexity, but also solves the problem that the number of iterations in the traditional model is difficult to be determined.