Whole-Heart CT Image Segmentation Based On Encoding-Decoding Structure

Cardiovascular and cerebrovascular diseases are the leading cause of death in human diseases around the world.For this reason,people invest a lot of energy to prevent and treat cardiovascular and cerebrovascular diseases.The pathological information and anatomical information provided by 3D medical imaging techniques such as CT images are doctors' diagnosis of cardiovascular and cerebrovascular diseases.One of the main reasons,such as the acquisition of the whole heart structure by CT images,plays an important role in the diagnosis of heart failure,congenital heart failure and guiding radiofrequency ablation surgery.As manual manual analysis of medical images is time-consuming and laborious,it is of great significance to quickly and accurately automate segmentation of medical images.This paper has studied the whole heart automated segmentation algorithm of cardiac CT images which are widely used in medical images.The first part of this paper first summarizes the current popular heart segmentation algorithm,and studies the two major difficulties in the current whole heart segmentation: the background tissue(that is,the tissue other than the heart substructure)has far more voxels than the heart substructures.The values of voxels between the boundaries of heart substructures,the boundaries of the heart substructure and the background tissue are very close,resulting in very blurred boundaries.After comparing the advantages and disadvantages of 2D segmentation and 3D segmentation,this paper uses the 3D UNET that best preserves the context information between slices as the infrastructure,and applies the deeply-supervised mechanism to improve the performance of the neural network on the small sample set.A multi-depth fusion block is proposed to assist the network fusion context information.Finally,Focal loss and Dice loss are combined to form a new loss function to suppress the class imbalance and reduce the loss of easy-divided voxels,forcing the network to optimize the loss of difficult voxels and improve the accuracy of the segmentation algorithm.The second part of this paper continues the research on the basis of the first part.Firstly,we studied the most advanced two-stage segmentation method in the whole heart segmentation field.The two-stage segmentation method first uses the positioning network to determine the approximate location of the heart and then completes the segmentation task based on the located region.On this basis,in order to achieve the same focus on the key areas where the heart substructure is located,the experiment used a multi-attention mechanism.Then,based on the fact that most background pixels gradually become non-information during training,a difficult negative sample mining method is used to adaptively exclude the loss of information-free background pixels from the loss function.Finally,this paper studies CliqueNet with a round-robin mechanism and feedback mechanism,and integrates Clique blocks into our network.The effect of each method was evaluated at the end of each section.In the end,the average Dice index of this experiment reached 91.01%.Compared with the results of other recent papers,the method proposed in this paper has reached the most advanced level.The method used in this experiment is end-to-end and can perform high-precision whole heart segmentation tasks in a relatively short period of time,which has a high significance for the clinical diagnosis of heart disease.This method is versatile and can therefore be migrated to a segmentation task similar to medical images such as brain CT images.