Ultrasound image segmentation of left ventricle is very important for doctors in clinical practice. As the ultrasound images contain a lot of noise and the contour features are not obvious, current Convolutional Neural Network (CNN) method is easy to obtain unnecessary regions in left ventricular segmentation, and the segmentation regions are incomplete. In order to solve these problems, keypoint location and image convex hull method were used to optimize segmentation results based on Fully Convolutional neural Network (FCN). Firstly, FCN was used to obtain preliminary segmentation results. Then, in order to remove erroneous regions in segmentation results, a CNN was proposed to locate three keypoints of left ventricle, by which erroneous regions were filtered out. Finally, in order to ensure that the remained area were able to be a complete ventricle, image convex hull algorithm was used to merge all the effective areas together. The experimental results show that the proposed method can greatly improve left ventricular segmentation results of ultrasound images based on FCN. Under the evaluation standard, the accuracy of results obtained by this method can be increased by nearly 15% compared with traditional CNN method.

Due to redundancy of ultrasound image samples, high similarity between different planes caused by disease, and inaccurate positioning of region-of-interest, a classification method of TransEsophageal Echocardiography (TEE) standard plane was proposed by combining with Bag of Features (BOF) model, active learning and improved multi-class AdaBoost algorithm. Firstly, BOF model was constructed to describe ultrasound image. Secondly, active learning was adopted to select the most informative samples for classifiers as training data set. Lastly, improved multi-class AdaBoost algorithm was proposed, where the weight update rule of multi-class AdaBoost was modified according to the classfication results of temporary strong learner, and the TEE standard plane was classified by the improved multi-class AdaBoost algorithm. The experimental results on TEE data set and three UCI data sets showed that, compared with AdaBoost.SAMME, multi-class Support Vector Machine (SVM), BP neural network and AdaBoost.M2, the G-mean value, the total classification accuracy and the classification accuracy in most classes of the proposed method were improved in varying degrees, the classification accuracy of easily misclassified class was improved most significantly. The experimental results illustrate that the improved multi-class AdaBoost algorithm can significantly improve the G-mean value and accuracy of easily misclassified class in the datasets containing similar samples between classes.

An imbalance multi-class AdaBoost algorithm with active learning was proposed to improve the recognition accuracy of minority class in imbalance classification. Firstly, active learning was adopted to select the most informative samples for classifiers through multiple iterations of sampling. Secondly, a new sample selection strategy based on uncertainty of dynamic margin was proposed to tackle the problem of data imbalance in the multi-class case. Finally, the cost sensitive method was adopted to improve the multi-class AdaBoost algorithm: giving different class with different misclassification cost, adjusting sample weight update speed, and forcing weak learners to "concern" minority class. The experimental results on clinical TransThoracic Echocardiography (TTE) data set illustrate that, when compared with multi-class Support Vector Machine (SVM), the total recognition accuracy of heart disease increases by 5.9%, G-mean improves by 18.2%, the recognition accuracy of Valvular Heart Disease (VHD) improves by 0.8%, the recognition accuracy of Infective Endocarditis (IE) (minority class) improves by 12.7% and the recognition accuracy of Coronary Artery Disease (CAD) (minority class) improves by 79.73%; compared with SMOTE-Boost, the total recognition accuracy of heart disease increases by 6.11%, the G-mean improves by 0.64%, the recognition accuracy of VHD improves by 11.07%, the recognition accuracy of Congenital Heart Disease (CHD) improves by 3.67%. The experiment results on TTE data and 4 UCI data sets illustrate that when used in imbalanced multi-class classification, the proposed algorithm can improve the recognition accuracy of minority class effectively, and upgrade the overall classifier performance while guaranteeing the recognition accuracy of other classes not to be decreased dramatically.

A deep model for automatic recognition of echocardiographic standard views based on deep convolutional neural network was proposed, and the effectiveness of the deep model was analyzed by visualize class activation maps. In order to overcome the shortcomings of the fully connected layer occupying most of the parameters of the model, the spatial pyramid mean pool was used to replace the fully connected layer, and more spatial structure information was obtained. The model parameters and the over-fitting risk were reduced.The attention mechanism was introduced into the model visualization process by the class significance region. The robustness and effectiveness of the deep convolution neural network model were explained by the case recognizing echocardiographic standard views. Visualization analysis on echocardiography show that the decision basis made by the improved depth model is consistent with the standard view classification by the sonographer which indicates the validity and practicability of the proposed method.

To solve the multi-label classification problem that a target belongs to multiple classes, a new multi-label classification algorithm based on floating threshold classifiers combination was proposed. Firstly, the theory and error estimation of the AdaBoost algorithm with floating threshold (AdaBoost.FT) were analyzed and discussed, and it was proved that AdaBoost.FT algorithm could overcome the defect of unstabitily when the fixed segmentation threshold classifier was used to classify the points near classifying boundary, the classification accuracy of single-label classification algorithm was improved. And then, the Binary Relevance (BR) method was introduced to apply AdaBoost.FT algorithm into multi-label classification problem, and the multi-label classification algorithm based on floating threshold classifiers combination was presented, namely multi-label AdaBoost.FT. The experimental results show that the average precision of multi-label AdaBoost. FT outperforms the other three multi-label algorithms, AdaBoost.MH (multiclass, multi-label version of AdaBoost based on Hamming loss), ML-kNN (Multi-Label k-Nearest Neighbor), RankSVM (Ranking Support Vector Machine) about 4%, 8%, 11% respectively in Emotions dataset, and is just little worse than RankSVM about 3%, 1% respectively in Scene and Yeast datasets. The experimental analyses show that multi-label AdaBoost. FT can obtain the better classification results in the datasets which have small number of labels or whose different labels are irrelevant.

It is called labels matching problem when two labels of an instance come from two labelsets respectively in multi-label classification, however there is no any specific algorithm for solving such problem. Although the labels matching problem could be solved by tranditional multi-label classification algorithms, but this problem has its own particularity. After analyzing the labels matching problem, a new labels matching algorithm based on pairwise labelsets was proposed using adaptive method, which considered the real Adaptive Boosting (real AdaBoost) and the global optimization idea. This algorithm could learn the rule of labels matching well and complete matching. The experimental results show that, compared with the traditional algorithms, the new algorithm can not only reduce searching scope of the labels space, but also decrease the minimum learning error as the number of weak classifiers increases, and make the classification more accurate and faster.

Medical image filtering process should retain the edge details of diagnostic significance. For Perona-Malik (PM) anisotropic diffusion model experienced failure when dealing with strong noise and choosing parameter K of diffusion threshold relies on experience, this paper proposed an improved anisotropic diffusion algorithm. First, PM was combined with the median filter algorithm, and then the gradient mode of the original image was replaced with the gradient mode from the image which was smoothed by the median filter to control the process of diffusion. While applying the adaptive diffusion threshold (Median Absolute Deviation (MAD) of the gradient in current neighborhood) and iteration termination criteria, the algorithm improved robustness and efficiency of the algorithm. The experiment was operated respectively on echocardiography, CT images and Lena image to denoise, and used Peak Signal-to-Noise Ratio (PSNR) and Edge Preservation Index (EPI) as evaluation criterion. The experimental results show that the improves algorithm outperforms PM algorithm and Catte-PM method for improving PSNR while preserving image detail information, and meets the requirements for application in medical images more effectively.

Since the anisotropic diffusion methods cannot make a distinction between strong noise and weak edge effectively, the authors proposed an improved anisotropic diffusion denoising method based on local statistical characteristics. While denoising images by anisotropic diffusion method, points with large gray-level variations were found by using gradient threshold, and whether the point was a noise point or not was judged by calculating local variance and local deleted variance, and then mean filtering was used for the noise points. The experiments upon simulation images and clinical ultrasonic images show that this method preserves features and edges more efficiently than traditional anisotropic diffusion methods while denoising images.