Research On Medical Image Segmentation Algorithm For Cardiovascular Diseases

Cardiovascular Diseases(CVD)are the most important disease burden in the world.It is very important to improve the prevention and treatment of CVD.Medical image seg-mentation is one of the key tasks in CVD assisted diagnosis and research,which aims to automatically identify the pixel region of the target from medical images.Medical image segmentation technology based on deep learning can quickly process a large amount of data,which is helpful for the auxiliary diagnosis and research of cardiovascular diseases.In this paper,a large number of complete annotated data are scarce and medical sequence data processing is difficult in the task of medical image segmentation for cardiovascular diseases.The main contents are divided into three aspects:1.Aiming at the problem of scarce labeled data in medical images,this paper pro-poses a semi-supervised medical image segmentation algorithm based on multi-scale pro-totype learning,which uses a large amount of unlabeled data and a small amount of labeled data to learn.Specifically,unlabeled data and rare labeled data are used for prototype learning on multi-scale features,so that labels of labeled data can guide the segmentation of unlabeled data through the prototype.In addition,in order to better obtain the relevant prior knowledge from the labeled information,the relational aggregation module is used to transfer the relevant information between the labeled data and the unlabeled data.Quanti-tative and qualitative evaluation experiments on MRI datasets for cardiac function screen-ing and CT data sets for abdominal fat risk assessment show that the proposed method is a versatile and effective semi-supervised medical image segmentation algorithm.2.In ultrasound carotid plaque scan sequences,it is very difficult to obtain a com-plete pixel annotation of the target area.Therefore,this paper proposes a weakly su-pervised learning-based carotid plaque segmentation algorithm for learning using more readily available scribble-annotated data.Given the scribble annotations of the first frame as a reference,the task of segmenting plaques in subsequent frames of a carotid ultra-sound video is implemented.First,use the reference annotation or the prediction result of the previous frame to extract the template and the search area of the current frame,use the Siamese network to extract the features,and realize the matching between the tem-plate area of the previous frame and the search area of the current frame.In addition,the template features are modified by combining the features of the first frame to alleviate the tracking drift problem in the Siamese network.Using the positioning information of the scribble annotation,the response position of the template to the search area is opti-mized through the scribble response loss.Experiments on the carotid plaque ultrasound video dataset show that our method can accurately identify and segment carotid plaques in ultrasound visual scan sequences.3.Due to the blurred boundaries and irregular motion of cardiac structures in echocar-diography,which brings challenges to the accurate segmentation of cardiac structures,this paper proposes an echocardiographic segmentation algorithm based on temporal feature matching,given the reference annotation of the first frame In the case of post-sequence frames,automatic segmentation of the left ventricle and myocardial regions is done.Specif-ically,firstly,the feature extraction of multi-scale context is performed on the echocardio-graphic video sequence to obtain spatial detail information and semantic features.On the other hand,combined with the reference annotation and the prediction result of the pre-vious frame,learn the semantic features of the target foreground pixels of the first frame and the previous frame,and then use the feature distance to match with the current frame to achieve information transfer,thereby guiding the segmentation of the current frame..Comparison and ablation experiments on the echocardiographic sequence video dataset demonstrate the effectiveness of the proposed method.