Segmentation And Quantification Of Multi-view Echocardiographic Sequences

Cardiac disease is currently one of the most important causes of death.Ultrasound is widely used in the diagnosis of cardiac disease,and the segmentation and quantification of multi-view echocardiographic sequences play a key role.With the rapid development of computing power,deep learning methods have begun to show outstanding application potential in the field of medical image analysis.However,the segmentation and quantification of multi-view echocardiographic sequences still face great challenges due mainly to the limited pixel-wise manual annotated data,the disturbance of noise,motion artifacts,and gaps in multi-view and multi-source data.Most of the existing methods only focus on the segmentation of a single view or specific frames,involving multi-step operations and taking a long time,which has great limitations.In this study,a total of 150 cases of heterogeneous data from various vendors in three hospitals were collected,containing 13500 images,and all images have pixel-wise annotation of left ventricle endocardium.To solve the aforementioned problems in the segmentation and quantification of multi-view echocardiographic sequences,this paper has proposed a dense feature pyramid and deep supervision network(DSN)and a recurrent aggregation learning double-branch network(RAN).First,taking into account the large amount of computation in temporal modeling and noise accumulation in recurrent neural networks when dealing with ultrasound data,DSN processed all frames independently without using temporal information,took full advantage of multi-level and multi-scale features and deep supervision mechanism to achieve precise segmentation and quantification;further,RAN tried to restrict the noise accumulation problem during temporal modeling.RAN made the beast of multi-level and multi-scale spatial-temporal features by a recurrent aggregation learning double-branch mechanism.Besides,Inspired by the characteristic of the dataset,RAN decoupled a classification branch as an auxiliary task.By recurrent aggregation of spatial-temporal features and the regularization from the classification branch,RAN gained prominent segmentation and quantification results as well as classification results.Adequate experiments(ablation,comparison,stability,noise,and clinic)showed the great generalization and robustness of DSN and RAN.Besides,the timing statistics also showed the real-time processing ability of DSN and RAN.All these reveal the clinical potential of DSN and RAN.