Research On Medical Image Segmentation And Classification For Computer Aided Diagnosis Of Cardiovascular Disease

Cardiovascular disease is the most important public health problem at present.Medical image plays an indispensable role in the computer aided diagnosis of cardiovascular disease.As a significant technology in the preliminary diagnosis of cardiovascular diseases,leucocyte differential count involves a complex process of leukocyte classification,and artificial classification is inefficient and subjective.As a important diagnostic method for a variety of cardiovascular diseases,cardiac magnetic resonance imaging needs a long time for a professional and experienced clinician to manually draw the tissue outline in the early stage.The developmen t of deep learning technology has brought new development and breakthrough to the automatic analysis of medical images.Using deep learning technology to solve the complex manual reading work and improve the efficiency and accuracy of clinicians in diagnosing cardiovascular diseases has become the key point of this paper.This paper studies the multi-classification of white blood cells on image data of Blood Cell Images.In view of the lack of original data and the unbalanced categories,the data enhancement method of generative adversarial network is adopted on the basis of the pre-data enhancement.This method can increase the diversity of the generated samples.On this basis,attention mechanism is introduced to learn global feature dependency.Based on the idea of “migration”,a semi-automated model self-increasing method is proposed to simulate the process of fine-tuning the network using different data,and better model parameters and higher classification accuracy are obtained under the same network and the same data.Through the contrast experiments,it is verified that the improved algorithm get better results in the classification accuracy and has the ability to achieve more accurate classification on unbalanced small data sets.In this paper,the research on automatic segmentation algorithm of cardiac multi-structure is carried out on the short-axis MR image data of automatic cardiac diagnosis challenge.Based on the mainstream medical image segmentation network U-Net,an improved neural network structure of dense mesh DNet Unet is proposed to improve feature multiplexing.The weighted Focal Loss is introduced for the problem that the segmentation categories are unbalanced and the right ventricular tissue is more difficult to segment.Taking the loss of the region as well as the loss of the contour into account,the Boundary Loss is proposed.A semi-supervised framework is proposed for a small amount of annotated data and a large amount of unlabeled data,and a semi-supervised framework with generative adversarial network is used to simplify the two-stage training task.In this paper,comparative experiments are designed to verify the effectiveness of various innovative optimization methods for cardiac MRI segmentation.And compared with the most advanced single model segmentation algorithm,the optimal results are obtained in a number of evaluation indexes.Finally,according to the prediction of the segmentation model proposed in this paper,we extract multiple features and use the fusion multiple classification models to verify the correctness of the segmentation results.