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.

At present, the research of face frontalization mainly solves the face yaw problem, and pays less attention to the face frontalization of the side face affected by yaw and pitch at the same time in real scenes such as surveillance video. Aiming at this problem and the problem of incomplete identity information retained in front face image generated by multi-angle side faces, a Generative Adversarial Network (GAN) based on feature map symmetry and periocular feature preserving loss was proposed. Firstly, according to the prior of face symmetry, a symmetry module of the feature map was proposed. The face key point detector was used to detect the position of nasal tip point, and mirror symmetry was performed to the feature map extracted by the encoder according to the nasal tip, so as to alleviate the lack of facial information at the feature level. Finally, benefiting from the idea of periocular recognition, the periocular feature preserving loss was added in the existing identity preserving method of generated image to train the generator to generate realistic and identity-preserving front face image. Experimental results show that the facial details of the images generated by the proposed algorithm were well preserved, and the average Rank-1 recognition rate of faces with all angles under the pitch of CAS-PEAL-R1 dataset is 99.03%, which can effectively solve the frontalization problem of multi-angle side faces.

Judging weather there exists mediastinal lymph node metastasis in the location of mediastinal lymph node region and correctly segmenting malignant lymph nodes have great significance to the diagnosis and treatment of lung cancer. In view of the large difference in mediastinal lymph node size, the imbalance of positive and negative samples and the feature similarity between surrounding soft tissues and lung tumors, a new cascaded two-stage mediastinal lymph node segmentation algorithm based on attention was proposed. First, a two-stage segmentation algorithm was designed based on the medical prior to remove mediastinal interference tissues and then segment the suspicious lymph nodes, so as to reduce the interference of the negative samples and the difficulty of training, while enhancing the ability to segment mediastinal lymph nodes. Second, a global aggregation module and a dual attention module were introduced to improve the network's ability to classify multi-scale targets and backgrounds. Experimental results showed that the proposed algorithm achieved an accuracy of 0.707 9, a recall of 0.726 9, and a Dice score of 0.701 1 on the mediastinal lymph node dataset. It can be seen that the proposed algorithm is significantly better than other current mediastinal lymph node segmentation algorithms in terms of accuracy and Dice score, and can solve problems such as the big difference in size, sample imbalance and easily confused features of lymph nodes.

Concerning the problem of tooth movement path planning in virtual orthodontic treatment system, a method of tooth movement path planning based on simplified mean particle swarm with normal distribution was proposed. Firstly, the mathematical models of single tooth and whole teeth were established. According to the characteristics of tooth movement, the orthodontic path planning problem was transformed into a constrained optimization problem. Secondly, based on the simplified particle swarm optimization algorithm, a Simplified Mean Particle Swarm Optimization based on the Normal distribution (NSMPSO) algorithm was proposed by introducing the idea of normal distribution and mean particle swarm optimization. Finally, a fitness function with high security was constructed from five aspects:translation path length, rotation angle, collision detection, single-stage tooth moving amount and rotation amount, so as to realize the orthodontic movement path planning. NSMPSO was compared with basic Particle Swarm Optimization (PSO) algorithm, the mean Particle Swarm Optimization (MPSO) algorithm and the Simplified Mean Particle Swarm Optimization with Dynamic adjustment of inertia weight(DSMPSO) algorithm. Results show that on Sphere, Griewank and Ackley, these three benchmark test functions, this improved algorithm tends to be stable and convergent within 50 iteration times, and has the fastest convergence speed and the highest convergence precision. Through the simulation experiments in Matlab, the optimal path obtained by the mathematical models and the improved algorithm is verified to be safe and reliable, which can provide assisted diagnosis for doctors.

In order to achieve more accurate diagnosis of benign and malignant of thyroid nodule ultrasound images and avoid unnecessary puncture or biopsy surgery, a feature combining method of conventional ultrasound imaging and ultrasound elastography based on Convolutional Neural Network (CNN) was proposed to improve the accuracy of benign and malignant classification of thyroid nodules. Firstly, large-scale natural image datasets were used by the convolutional network model for pre-training, and the feature parameters were transferred to the ultrasound image domain by transfer learning to generate depth features and process small samples. Then, the depth feature maps of conventional ultrasound imaging and ultrasound elastography were combined to form a hybrid feature space. Finally, the classification task was completed in the hybrid feature space, and an end-to-end convolution network model was constructed. The experiments were carried out on 1156 images, the proposed method had the accuracy of 0.924, which was higher than that of other single data source methods. The experimental results show that, the edge and texture features of the image are shared by the shallow convolutions, the abstract features of the high-level convolutions are related to the specific classification tasks, and the transfer learning method can solve the problem of insufficient data samples. At the same time, the elastic ultrasound image can objectively quantify the lesion hardness of thyroid nodules, and with the combination of the texture contour features of conventional ultrasound image, the mixed features can more fully describe the differences between different lesions. Therefore, this method can effectively and accurately classify the thyroid nodules, reduce the pain of patients, and provide doctors with more accurate auxiliary diagnostic information.

Aiming at the problem of low information extraction rate of traditional methods and the unideal coloring effect in the grayscale image colorization field, a grayscale image colorization algorithm based on dense neural network was proposed to improve the colorization effect and make the information of image be better observed by human eyes. With making full use of the high information extraction efficiency of dense neural network, an end-to-end deep learning model was built and trained to extract multiple types of information and features in the image. During the training, the loss of the network output result (such as information loss and classification loss) was gradually reduced by comparing with the original image. After the training, with only a grayscale image input into the trained network, a full and vibrant vivid color image was able to be obtained. The experimental results show that the introduction of dense network can effectively alleviate the problems such as color leakage, loss of detail information and low contrast, during the colorization process. The coloring effect has achieved significant improvement compared with the current advanced coloring methods based on Visual Geometry Group (VGG)-net, U-Net, dual stream network structure, Residual Network (ResNet), etc.

Aiming at the insufficiency of feature extraction in the existing Content-Based Medical Image Retrieval (CBMIR) algorithms, which resulted in imperfect semantic information representation and poor image retrieval performance, an algorithm based on multi-scale dense network was proposed. Firstly, the size of pulmonary nodule image was reduced from 512×512 to 64×64, and the dense block was added to solve the gap between the extracted low-level features and high-level semantic features. Secondly, as the information of pulmonary nodule images extracted by different layers in the network was varied, in order to improve the retrieval accuracy and efficiency, the multi-scale method was used to combine the global features of the image and the local features of the nodules, so as to generate the retrieval hash code. Finally, the experimental results show that compared with the Adaptive Bit Retrieval (ABR) algorithm, the average retrieval accuracy for pulmonary nodule images based on the proposed algorithm under 64-bit hash code length can reach 91.17%, which is increased by 3.5 percentage points; and the average time required to retrieve a lung slice is 48 μs. The retrieval results of the proposed algorithm are superior to other comparative network structures in expressing rich semantic features and retrieval efficiency of images. The proposed algorithm can contribute to doctor diagnosis and patient treament.

Aiming at the problem that the deep learning network model applied to medical image target detection only has a fixed receptive field and cannot effectively detect the cervical lymph nodes with obvious morphological and scale differences, a new recognition algorithm based on adaptive receptive field mechanism was proposed, applying deep learning to the automatic recognition of cervical lymph nodes in complete three-dimensional medical images at the first time. Firstly, the semi-random sampling method was used to crop the medical sequence images to generate the grid-based local image blocks and the corresponding truth labels. Then, the DeepNode network based on the adaptive receptive field mechanism was constructed and trained through the local image blocks and labels. Finally, the trained DeepNode network model was used for prediction. By inputting the whole sequence images, the cervical lymph node recognition results corresponding to the whole sequence was obtained end-to-end and quickly. On the cervical lymph node dataset, the cervical lymph node recognition using the DeepNode network has the recall rate of 98.13%, the precision of 97.38%, and the number of false positives per scan is only 29, and the time consumption is relatively shorter. The analysis of the experimental results shows that compared with current algorithms such as the combination of two-dimensional and three-dimensional convolutional neural networks, the general three-dimensional object detection and the weak supervised location based recognition, the proposed algorithm can realize the automatic recognition of cervical lymph nodes and obtain the best recognition results. The algorithm is end-to-end, simple and efficient, easy to be extended to three-dimensional target detection tasks for other medical images and can be applied to clinical diagnosis and treatment.

In order to solve the problem of indoor crowd target scale and attitude diversity and confusion of head targets with surrounding objects, a new Network based on Multi-level Features and hybrid Attention mechanism for indoor crowd detection (MFANet) was proposed. It is composed of three parts:feature fusion module, multi-scale dilated convolution pyramid feature decomposition module, and hybrid attention module. Firstly, by combining the information of shallow features and intermediate layer features, a fusion feature containing context information was formed to solve the problem of the lack of semantic information and the weakness of classification ability of the small targets in the shallow feature map. Then, with the characteristics of increasing the receptive field without increasing the parameters, the dilated convolution was used to perform the multi-scale decomposition on the fusion features to form a new small target detection branch, realizing the positioning and detection of the multi-scale targets by the network. Finally, the local fusion attention module was used to integrate the global pixel correlation space attention and channel attention to enhance the features with large contribution on the key information in order to improve the ability of distinguishing target from background. The experimental results show that the proposed method achieves an accuracy of 0.94, a recall rate of 0.91 and an F1 score of 0.92 on the indoor monitoring scene dataset SCUT-HEAD. All of these three are significantly better than those of other algorithms currently used for indoor crowd detection.

The existing automatic recognition algorithms for cervical lymph nodes have low efficiency, and the overall false positive removal are unsatisfied, so a cervical lymph node detection algorithm using cascaded Fully Convolutional Neural Networks (FCNs) was proposed. Firstly, combined with the prior knowledge of doctors, the cascaded FCNs were used for preliminary identification, that was, the first FCN was used for extracting the cervical lymph node region from the Computed Tomography (CT) image of head and neck. Then, the second FCN was used to extract the lymph node candidate samples from the region, and merging them at the three-dimensional (3D) level to generate a 3D image block. Finally, the proposed feature block average pooling method was introduced into the 3D classification network, and the 3D input image blocks with different scales were classified into two classes to remove false positives. On the cervical lymph node dataset, the recall of cervical lymph nodes identified by cascaded FCNs is up to 97.23%, the classification accuracy of the 3D classification network with feature block average pooling can achieve 98.7%. After removing false positives, the accuracy of final result reaches 93.26%. Experimental results show that the proposed algorithm can realize the automatic recognition of cervical lymph nodes with high recall and accuracy, which is better than the current methods reported in the literatures; it is simple and efficient, easy to extend to other tasks of 3D medical images recognition.