Research On Coronary Artery Segmentation Based On Deep Learning

Cardiovascular disease（CVD）is one of the major health problems in the world today and most cardiovascular diseases can be attributed to problems with coronary artery stenosis.Computed tomography angiography（CTA）is used in the diagnosis and treatment of coronary artery disease due to its high resolution and non-invasive nature.Accurate coronary artery segmentation plays an important role in the diagnosis and treatment of coronary artery lesions.However,segmentation methods that require manual intervention are struggling to cope with the increasing volume of data,and fully automated segmentation methods are becoming increasingly important in coronary artery segmentation.In recent years,Convolution neural networks（CNN）have been widely used in deep learning for the automatic segmentation of medical images.Due to the high resolution and high dimensionality of CTA images,segmentation of coronary arteries using CNN on CTA images requires down-sampling or conversion of the images to slices and patches for segmentation to reduce computational resources.However,down-sampling results in the loss of local detail in the image,while slices and patches contain only local information about the image and lack global information,which limits the segmentation performance of the model.To address the above issues,this thesis investigates deep learning methods for segmenting coronary arteries based on CTA image data,of which the research is as follows.（1）In order to effectively fuse the global and local information of an image for segmentation,this thesis proposes an ensemble segmentation model based on U-Net.The model learns global and local information of CTA images by constructing a lightweight 3D U-Net segmentation model,a 2D U-Net multi-slice segmentation model and a 3D U-Net patch segmentation model,and finally integrating each model to obtain segmentation results.Experiments show that the U-Net-based ensemble segmentation model has a higher accuracy rate compared to single direct segmentation as well as slice and patch segmentation.（2）Although the U-Net-based ensemble segmentation model achieves good accuracy for segmenting coronary arteries,it requires training multiple segmentation models at high resolution and has a high overall computational overhead.Therefore,this thesis proposes a two-stage segmentation model based on dilated prior region that guides local image patches for segmentation by learning global dilated prior.The model first extracts the dilated global prior region at low resolution via 3D U-Net,divides the CTA image into patches based on the prior region and uses 3D U-Net++ for segmentation.Compared to the ensemble segmentation model of U-Net,this method has improved accuracy in coronary segmentation and better segmentation results using dilated priori regions than non-dilated priori regions.（3）In order to further improve the accuracy of coronary artery segmentation as well as to improve the stability of the segmentation.In this thesis,an ensemble segmentation model based on 3D Triplet attention and multi-scale is proposed.The method improves the accuracy of patch segmentation by adding the 3D Triplet attention mechanism to 3D U-Net++ and as patch segmentation network with an ensemble segmentation framework.On the other hand,the 3D U-Net++ network was trained by making patches of different sizes from dilated prior regions and non-dilated prior regions to obtain multiple local scale segmentations,while adding the global scale segmentation of the lightweight 3D U-Net.The experimental results show that the model has a higher segmentation accuracy compared to other segmentation models.