MRI-based Methods For The Biventricular Segmentation And Registration

Cardiovascular disease(CVD)is one of the most threatening diseases to the health of human life.Abnormal ventricular deformation significantly affects cardiac physiological status and leads to cardiovascular disease,and accurate and non-invasive quantitative analysis of the ventricle is important for CVD diagnosis.With the development of deep learning technology,computer-aided diagnosis(CAD)technology is an important contribution to the improvement of the quality and efficiency of cardiac disease diagnosis.Cardiac magnetic resonance images(MRI)is an effective tool for assessing ventricular function.Although some progress has been made in cardiac image segmentation and registration based deep learning,the existing segmentation models do not focus enough on small and medium target regions of the heart,and there are still broken and mis-segmented regions after segmentation.The segmentation performance of the model needs to be improved.In addition,in the registration task,the cardiac dataset often lacks the gold standard for the registration task,and there are a large number of redundant features leading to a huge number of computational parameters and poor registration performance.In this paper,the above problems are studied respectively and the relevant cardiac functional parameters are calculated based on the above segmentation and registration results.The main research contents of this paper are as follows:(1)In this paper,a Trans-ConvNet structure DPiTNet(Deformable Patch embedding and Token Pooling-based Vision Transformer Network)is proposed for cardiac image segmentation,which is based on the UNet-style encoder-decoder structure as the backbone.The 2D branch mainly consists of a DPiT encoder consisting of a deformable patch embedding module and a multi-scale token pooling module.2D-3D dimensional fusion modules and the multi-scale convolutional kernel fusion module are included in 3D branch.The experimental results show that the DPiT encoder module,the 2D-3D dimensional fusion module,and the multiscale convolutional kernel fusion module can all contribute to the improvement of the model performance,and after fusing the three modules,our model can achieve the optimal performance,which exceeds most of the existing mainstream segmentation algorithms.(2)In this paper,a two-stage registration pipeline named MAE-Trans RNet(Transformerbased Registration Network with Masked Autoencoder)is proposed based on the Trans-ConvNet structure.The masked autoencoder is used to learn more advanced abstract features.Depth Wise Separable Convolution(DWSC)and Squeeze and Excitation(SE)modules are introduced into the improved Multi-Head Self-Attention mechanism in the Transformer encoder,in order to improve the query-key-value-based dot product attention,and further reduce the amount of parameter computation.In addition,concurrent spatial and channel squeeze & excitation(scSE)module is embedded into the CNN structure,which also proves to be effective for extracting robust feature representations.The results of ablation experiments and comparison experiments validate the effectiveness of the improved module and the progressiveness of MAE-TransRNet.(3)Based on above segmentation and registration results,additional parameter information such as height,weight,and disease classification in the cardiac images is introduced to assist in the initial qualitative and quantitative functional evaluation of the heart by calculating the patient’s ventricular volume,ejection fraction,myocardial mass,and ventricular wall thickness closely related to heart disease.