The Analysis And Research Of Cardiovascular Medical Image Based On Deep Learning

With the popularization of computers and the rapid development of the IT industry,the scale of application systems in various industries has continued to expand,and the data generated has also exploded.At present,there is a large amount of unstructured data accumulation in the medical industry.Therefore,it is of great significance to study how to effectively use medical image data to improve the diagnostic efficiency.In recent years,our country actively promotes the development of medical informatization.The application of cutting-edge technologies such as big data and artificial intelligence in the medical field becomes a new trend.We focus on the cardiovascular field in this thesis,introduce the relevant deep learning methods for the automated analysis and diagnosis of medical imaging,and finally demonstrate the feasibility of AI auxiliary medical care.Our main mission is to complete the overall diagnosis of cardiovascular medical imaging and the location of vulnerable lesions.In this thesis,we study the classical object detection and semantic segmentation algorithms based on convolutional neural networks and analyze their applications in medical imaging.Secondly,according to the characteristics of our data and mission,we design a new network structure containing a classification branch and a regression branch.Through multi-task joint training,both the diagnosis and the detection effects were improved.For the unbalanced problem between classes in medical image analysis,this thesis also introduces a weighted loss function mechanism,which further enhanced the model effect.In reality,there are a large number of medical images with high-level category label information,lacking low-level and pixel-level label information.Unlike general images,medical image requires specialized experts to label,so large-scale medical data labeling is very difficult.To solve the problem,we use the cardiovascular imaging data as an example to study the possibility of lesion localization under weak supervision condition in this thesis.The channel weight allocation mechanism is introduced to apply the CAM(Class Activation Map)idea to different resolutions,and we compare the localization effects of different layers under weak supervision.At the same time,the network is constrained by multiple embedded spatial region weight learning modules.Finally,the effect of lesion localization under weak supervision is further enhanced.