Accurate cross-modality cardiac image segmentation is essential for effectively diagnosing and treating heart disease. Different imaging modalities help to determine suitable pre-procedure planning. However, most methods face the difficulty of spatial-temporal confounding, where the anatomy element and modality element of cardiac images are intertwined across both spatial and temporal dimensions. It is derived from the imaging diversity and structure diversity of cardiac images. The spatial-temporal confounding hinders knowledge transfer between cardiac images on different modalities. In this paper, we propose a novel dynamic causal learning (DCL) to solve spatial-temporal confounding. The DCL explores multi-dimensional causal intervention to consider not only the causal relationship between images and labels, but also the causality in time dimension and space dimension. It integrates historical optimal interventions and facilitates the transfer of this knowledge across temporal contexts. In addition, the DCL utilizes the diffusion mechanism to further ensure that the extracted anatomy element remains causal invariant, improving model performance across multiple imaging modalities. Extensive experiments on cross-modality cardiac images (MR, CT, and US) demonstrate the effectiveness of the DCL (mean Dice = 0.951), outperforming other advanced segmentation methods. DCL is freely accessible at https://github.com/asdww0721ww/DCL.

Echocardiography video segmentation is critical for cardiovascular disease diagnosis. However, it still suffers from the challenge of dual-level bias. This challenge derives from the frame-level bias in temporal dimension and the object-level bias in the spatial dimension on echocardiography video. To overcome this challenge, we propose a spatial-temporal consistency (STC) model based on semi-supervised learning for echocardiography video segmentation. This model aligns and fuses inter-frame and inter-object context-aware feature representations. First, the STC explores a temporal context-aware (TCA) module to focus on motion differences between frames. This module extracts temporal correlation through inter-frame attention to compensate for important temporal semantic information. Second, the STC proposes a multi-object semantic adaptation (MSA) module that not only adaptively calibrates frame-level feature and object-level feature, but also fuses these features at different layers. Finally, the STC considers spatial-temporal consistency constraint to reduce prediction error among multiple MSA modules, thereby achieving low-entropy prediction. Extensive experiments demonstrate that the STC achieves state-of-the-art performance for echocardiography video segmentation.

We introduce MorphMamba, a new deep learning architecture for 3D medical image segmentation that successfully adapts the efficiency of state space models to volumetric data. While standard Mamba models excel at long-sequence modeling, they fail to preserve the rich 3D spatial context of medical images due to a required data-flattening step. MorphMamba overcomes this critical limitation with a novel Mamba-based encoder designed for volumetric inputs. Our architecture integrates two key innovations: a Channel-Spatial Attention Guided module, which refines feature maps to emphasize salient anatomical information before sequential processing, and a Context Refining Block in the skip-connections to ensure optimal multi-scale feature fusion. We demonstrate the state-of-the-art performance of MorphMamba through extensive experiments on four diverse benchmarks, including BraTS 2023, LiTS 2017, a private cardiac dataset and a private prostate tumor dataset. Our model consistently achieves superior results in terms of both accuracy and boundary delineation, establishing a new and efficient baseline for 3D segmentation. Our code is available at https://github.com/haoranGeng/MorphMamba.

Domain generalization aims to learn common knowledge from multiple observed source domains and transfer it to unseen target domains, e.g. the object recognition in varieties of visual environments. Traditional domain generalization methods aim to learn the feature representation of the raw data with its distribution invariant across domains. This relies on the assumption that the two posterior distributions (the distributions of the label given the feature distribution and given the raw data) are stable in different domains. However, this does not always hold in many practical situations. In this paper, we relax the above assumption by permitting the posterior distribution of the label given the raw data changes in difference domains, and thus focuses on a more realistic learning problem that infers the conditional domain-invariant feature representation. Specifically, a multi-domain adversarial variational Bayesian inference approach is proposed to minimize the inter-domain discrepancy of the conditional distributions of the feature given the label. Besides, it is imposed by the constraints from the adversarial learning and feedback mechanism to enhance the condition invariant feature representation. The extensive experiments on two datasets demonstrate the effectiveness of our approach, as well as the state-of-the-art performance comparing with thirteen methods.