Prior-Knowledge Based Feature Learning In Echocardiography Intelligent Analysis

Deep learning technologies have achieved great success in medical image analysis with high accuracy,efficiency,stability,and scalability.It is hopeful to build a computer-aided diagnosis method based on deep learning technology,help doctors to complete the task of medical image analysis more conveniently and accurately,and serve practical clinical applications as a diagnosis-aided tool.However,it is not the best solution to analyze medical images using the unified model in computer vision.Compared with natural images in computer vision,prior knowledge such as pixel intensity changes,anatomical structures,and imaging information in medical images may all be important cues.However,traditional data-driven deep learning models ignore such prior knowledge.Constructing a deep learning model combined with prior knowledge for medical image analysis,can not only improve the learning ability of the model but also meet the needs of diagnosis in clinical scenarios.Specifically,combining prior knowledge with deep learning for medical image analysis mainly faces two needs and challenges:1)How to learn prior knowledge in medical images and incorporate it into image analysis tasks.Choosing appropriate prior knowledge for different medical analysis tasks,and incorporating it into the model to improve its performance to learn the task-specific features is difficult.2)How to help doctors to understand whether the results are reliable.Assisting clinical diagnosis and decision-making to make the medical image analysis model provide information that can help doctors to judge the reliability of the results,is a key challenge.To address the above challenges,this dissertation studies knowledge-guided image feature learning methods of different tasks in echocardiographic analysis.Based on prior knowledge such as the anatomical knowledge in the labeled image,the same region for segmentation and quantification,and the spatial and spectral distribution of images from different ultrasound devices,different models of prior-knowledge-based feature learning are designed for echocardiography intelligent analysis.The main research work in this dissertation can be summarized as follows:Firstly,a multi-constrained aggregate learning for echocardiographic myocardial segmentation is proposed.It leverages anatomical knowledge learned through ground-truth labels to infer segmented parts and discriminate boundary pixels.The new framework encourages the model to focus on the features in accordance with the learned anatomical representations,and the training objectives incorporate a boundary distance transform weight(BDTW)to enforce a higher weight value on the boundary region,which helps to improve the segmentation accuracy.The comparison study shows that the proposed network outperforms the other segmentation baseline models,indicating that our method is beneficial for boundary pixel discrimination in segmentation.Secondly,a multi-task model based on task-relation spatial co-attention is proposed for paired 2D ultrasound echocardiogram segmentation,quantification,and uncertainty estimation.The method enables multi-task learning by innovatively exploring spatial correlations between tasks,by using non-local and co-attention to learn spatial mappings between task-specific features,thereby forcing the aggregation of spatial information embedded in segmentation and quantification tasks.Boundary-aware structural consistency and joint indices constraint are integrated into the multi-task learning optimization objective,to guide the learning of segmentation and quantization paths.The former constrains the predicted structural similarity,while the latter explores the intrinsic relationship among these three quantitative indices.At the same time,in order to provide the reliability of the prediction results,TRSA-Net innovatively designs a dual-branch structure to calculate the uncertainty of model segmentation results.This design leverages the inconsistency of the prediction from the two branches to constrain the segmentation results.Given the uncertainty of the model while obtaining high-accuracy segmentation results.Experiments on public datasets show that the proposed method can achieve competitive segmentation performance with highly accurate quantification results while providing the uncertainty of segmentation results.Finally,a unified bispace-constrained(spatial and spectral)domain adaptation for cross-domain echocardiography segmentation,which integrates style-regularized image disentanglement,bispace-constrained image translation,and confidence-aware semantic segmentation.Style-regularized image disentanglement maps the domain-specific style attributes into a predefined Gaussian distribution for diversity cross-domain image domain adaptation.Bispace-constrained image translation learns domain-invariant content by incorporating the 1D power spectral density and spatial spaces into the diversity image translation for robust domain-invariant feature extraction.Confidence-aware semantic segmentation leverages the aligned information from image translation to guide effective cross-domain segmentation.Results on two public datasets(CAMUS and EchoNet-Dynamic)for cardiac structure segmentation demonstrated that the proposed model has comparable performance to existing unsupervised domain adaptation segmentation methods in 2D echocardiography,thus can provide a new framework for efficient clinical application.