| Cardiovascular system is the circulatory system of the human body,which assumes the role of transporting nutrients and discharging metabolic waste.However,at present,cardiovascular disease is the number one health killer known to everyone and its morbidity and mortality show a trend of increasing year by year,so obtaining anatomical information about the cardiovascular system is crucial for the prevention and diagnosis and treatment of cardiovascular diseases,and interventional surgeries such as cardiac puncture also require a complete cardiovascular segmentation map.However,in CT images,the diameter and edge morphology of each structure of the whole-heart are quite different,and the gray scale change is small,which makes it impractical to manually segment the heart structure clinically.Traditional cardiac segmentation methods based on thresholds and clustering have problems such as low segmentation accuracy,while 2D and 3D segmentation networks based on deep learning also have problems such as accuracy to be improved,computation volume,and memory consumption.Therefore,aiming at the structural characteristics of the wholeheart CT images and the difficulty of automatic segmentation,this paper proposes the2.5D whole-heart segmentation algorithm based on deep learning.Firstly,U-Net＿conv1-5Mffp network based on parallel multi-scale feature fusion is proposed.The network adds parallel multi-scale feature fusion modules composed of different dilation rates to different layers of the U-Net network encoder,expands the receptor field and captures more information of interest,and to a certain extent improves the segmentation of small vessels of the pulmonary artery and atrium,ventricles and other structures,but there is still a certain lack of segmentation of the boundaries of each structure.Secondly,U-Net networks based on attention mechanisms are proposed.By distinguishing the importance of features and assigning different weights to different features,the attention mechanism makes the network pay more attention to the required key information.Therefore,the AG-U-Net network based on local and global attention computing models is proposed,which makes it pay more attention to boundary information and weaken background information,which effectively compensates for the loss of poor segmentation of the U-Net＿conv1-5Mffp network.Based on the advantages of parallel multi-scale feature fusion and attention mechanism,the AG-UNet＿conv1-5Mffp network is proposed,and cross entropy,which balances the proportion of positive and negative samples,is used as the loss function.The network not only captures more global information,but also highlights the heart boundary characteristics of interest,and to some extent solves the problem of imbalance between positive and negative samples.However,the network doesn’t consider threedimensional spatial information,and its segmentation effect still needs to be improved.Finally,based on the shortcomings of AG-U-Net＿conv1-5Mffp network,the 2.5D whole-heart segmentation model of adaptive information fusion is proposed.Based on the 2D segmentation results of the three sections of the heart CT images by the AG-UNet＿conv1-5Mffp network,the model uses the multi-view adaptive fusion mechanism to learn,assigns different weights to the same pixel of different sections,and fuses the multi-view segmentation information,obtaining more comprehensive and fine threedimensional spatial information of the heart,and improving the accuracy of the segmentation algorithm.To verify the performance of the proposed whole-heart segmentation network,the model was trained and evaluated on the MM-WHS dataset.Experimental results show that the Dice value of the 2.5D whole-heart segmentation model proposed in this paper reaches 91.45%,which effectively verifies the performance of the algorithm,and the algorithm reduces the computing resources to a certain extent,which is expected to be integrated into the computer-aided diagnosis system to meet clinical needs. |
PDF Full Text Request