Design And Implementation Of Ventricular Segmentation System Based On Deep Learning

As the first threat to human health,cardiovascular disease gradually becomes more and more serious with the improvement of living standars.Early detection and timely and accurate treatment are important for saving the lives of patients.The continuous improvement of medical imaging technology and the further enrichment of acquisition equipment have led to a significant increase in the number and quality of medical images.Medical image segmentation serves as a basis for other medical image processing and analysis tasks.The quality of the performance is directly related to the quality of subsequent tasks.In this paper,I use deep-learning techniques developed rapidly in the field of computer vision in recent years to conduct in-depth studies on the two tasks of ventricular segmentation and left ventricular scar tissue segmentation.For the ventricular segmentation problem,this paper first crops the original image according to the statistical information of the data and extracts the region of interest.For the ventricular segmentation study,the fully convolutional neural network widely used in natural image segmentation is used for segmentation.After that,the Conv LSTM structure is added to the network to form an improved network.At last,the densely conneted conditional random field algorithm is used to postprocess the segmentation results to obtain the final segmentation result.For the original network and the improved network,experiments and analyses were conducted separately.For the left ventricular scar segmentation task,this paper first crops and fills the original image according to the label of the left ventricle.After preprocessing,all images have same size.In the study of scar segmentation,The U-Net is used as our basic structure.Then,the part of feature extrction in the U-Net is replaced by a dense connection block structure,and the original downsampling process is replaced by the transition layer to form an improved Dense U-Net structure.In addition,this paper designs a network to to add the dense connection conditions of th form of a recurrent neural network to the U-Net network structure.Finally,this paper combines these two structures and U-Net into the same network and gets our fourth network structure.In the experiment,the performance of these four network structures and the two additional methods of segmentation and full width at half maximum were evaluated on multiple indexes.Based on the above deep learning algorithm,this paper uses Py Qt to design and implement the ventricular segmentation system in Windows system.After detailed functional tests,the system can use the neural network designed in this paper to perform efficient and accurate segmentation,and can also evaluate the performance of segmentation results.The system of this paper has high usability and expansibility and can provide convenience for users.