Cardiac Late Gadolinium Enhanced Image Segmentation Algorithm And System Based On Domain Adaptation

Late Gadolinium Enhancement(LGE)technology is a kind of cardiac magnetic resonance(CMR)imaging technology and is an important basis for the diagnosis of myocardial infarction.Segmenting the lesion area of the cardiac LGE image and obtaining relevant quantitative indicators is usually an important part of the clinical diagnosis of myocardial infarction.However,it takes a lot of time to manually delineate the corresponding structures in the heart.Deep learning is fast,reproducible,and safe,which is just suitable for solving the problem of cardiac LGE image segmentation.With the rapid development of deep learning technology,more and more deep learning-based cardiac LGE image segmentation methods have been introduced and achieved certain results,but there are still some problems to be solved.Including,but not limited to,the spatial structure and position of lesions in cardiac LGE images are not fixed,and it is difficult to perceive the complete spatial structure.The proportion of lesion areas is small,which leads to poor segmentation results.At the same time,the acquisition of cardiac LGE images is difficult,resulting in a small amount of data and insufficient high-quality labels.In response to the above problems,this thesis has carried out the following work:(1)In this thesis,a spatial location-aware cardiac LGE image lesion segmentation network(SLANet)is designed to deal with the problem that the spatial location of the lesion area is not fixed and the proportion is small in cardiac LGE image segmentation.The network structure of SLANet is designed based on the inherent spatial position relationship of the heart tissue,and at the same time,a small structure attention module is embedded,which helps the network to learn the characteristics of small lesion areas and improve the segmentation accuracy of the lesion area.A joint training scheme is designed to jointly train the segmentation network to improve the segmentation performance,and finally achieve the segmentation of normal physiological tissues and lesion areas in the heart.Experiments show that SLANet achieves better segmentation results than some methods on different cardiac LGE image datasets.(2)Based on SLANet,this thesis designs a domain-adaptive cardiac LGE image segmentation framework(DASeg)to deal with the generally small amount of cardiac LGE image data and improve the segmentation accuracy of lesion area.DASeg includes two modules:domain adaptation module and segmentation module.The domain adaptation module achieves the registration of cardiac LGE images and cine CMR images between different patients through a cross-modality recurrent registration network CMCRNet.The segmentation module uses SLANet.DASeg introduces the feature information of cine CMR image data into cardiac LGE images to achieve domain adaptation,assists the training of cardiac LGE image segmentation,and solves the problem that the amount of cardiac LGE image data is generally small to a certain extent.Experiments have verified that the segmentation accuracy of DASeg is further improved compared to SLANet on the same dataset.(3)In order to realize the clinical application of the above segmentation algorithm,this thesis designs and develops a cardiac LGE image segmentation processing system LGE Seg.The design of LGE Seg is based on the view/model/controller mode and is implemented using the PyQt5 toolkit.It is a lightweight PC software with functions such as cardiac LGE image display,image preprocessing,deep learning segmentation model calling and segmentation index calculation.