| Currently,medical imaging technologies used for cardiovascular diseases include magnetic resonance imaging(MRI)and computed tomography(CT).Doctors can use these digital images for cardiac image segmentation,functional parameter quantification,and partial pathological identification,leading to quick and accurate cardiovascular disease diagnoses.However,faced with a large amount of multi-modality cardiac images,manual segmentation operations based on doctors’ experience are no longer sufficient to meet clinical needs.Moreover,existing deep learning-based methods cannot deal with the four challenges of multi-modality cardiac images:images with multi-modalities and dense pixel resolution;sparse and noisy labels;imbalanced samples with multi-modal distributions;heterogeneous and isolated data.Therefore,this thesis proposes a deep learning-based algorithm for multi-modality cardiac image segmentation under supervised,unsupervised and self-supervised scenarios.Specifically,the main contributions of this thesis are as follow:To solve the challenge that images with multi-modalities and dense pixel resolution,this thesis proposes a deep neural network based on multiscale deformable convolutions for supervised cardiac MRI short-axis cine image segmentation.The method first proposes a joint offset prediction network to extract spatio-temporal information from consecutive cardiac MRI slices,including a target slice and its neighboring reference slices,which could reduce the adverse effects of cardiac motion artifacts.The method also proposes an enhanced deformable attention network,with a pyramidal and cascading architecture,to generate features with different scales by several flexible deformable convolutional layers.Then features in lower scales are aligned with coarse estimations,and the deformable offsets and aligned features are propagated to higher scales for precise cardiac MRI segmentation.In addition,MSAM is proposed to embrace external statistic and capture long range dependencies between different scale features.The method finally adopts a two-stream CNN with shared weights,to model the feature as a distribution,which enhances robustness and prevents overfitting.To solve the challenge that sparse and noisy labels and imbalanced samples with multimodal distributions,this thesis proposes a deep learning method based on a mixed high-order attention network for supervised cardiac CT image segmentation.The method estimates and maximizes the mutual information between input data and learned high-level representations,which pushes the model to learn the discriminative and compact features and reduces the negative influence of noisy images.The method also designs mixed high-order attention network to focus on important information,increase receptive fields,and suppress the expression of irrelevant regions.The method finally proposes a multi-expert architecture to generate segmentation results and provide confidence of results.The experimental results show that the proposed method achieves higher performance than other existing methods.To solve the challenge that heterogeneous and isolated data,this thesis proposes a deep neural network based on partial unbalanced feature transport for unsupervised cross-modality cardiac image segmentation.The probabilistic posterior distributions of latent features are approximated directly by using inverse continuous normalizing flows,which is beneficial to learning a shared geometrical and continuous latent space for both the source and target domains.The method further proposes a partial unbalanced optimal transport strategy,which relaxes the marginal constraints of the unbalanced total probability masses.It does not follow the oneto-one matching principle,and leads to partial transport and accurate domain alignment.The method also derives the Fenchel-Legendre dual form of PUOT for efficient calculation and the deviation bounds of PUOT for error analysis.Experimental results show that the proposed method achieves higher segmentation performance compared to existing unsupervised domain adaptation methods.To solve the challenge that heterogeneous and isolated data in the source-free condition,this thesis proposes a deep neural network based on structural mutual information for selfsupervised cross-modality cardiac image segmentation.The method adopts an unsupervised learning framework of iterative clustering,which uses prior knowledge from the pre-trained model to cluster the pixels of input images for preliminary segmentation.Then,the method proposes a prototype feature-based clustering method,which combines high-dimensional feature invariance based on photometric transformations and high-dimensional feature equivariance based on geometric transformations.It could achieve self-supervised learning under cross views and improve the model’s ability to segment local and global pixels.The method finally proposes a structural mutual information estimation strategy,which can enhance the strong dependence between pixel points in cardiac image by maximizing the structural mutual information between the segmented images and pseudo label under cross view.It also could overcome the shortcomings of the per-pixel loss function,which always ignores structural information.The effectiveness of the method is verified in several experiments. |
