Abstract:

[Purpose] The data distribution has a significant impact on the performance of deep learning models. However, deep models may learn features irrelevant to the segmentation target, which is usually inapplicable for new datasets, resulting in insufficient generalization ability. [Methods] To alleviate this problem, this paper proposes a Transformer-based medical image segmentation method with causal restraint. Taking MCRformer as the main body of the network, the Morphological Constraint Stream module is used to extract morphological constraint prior information. The meshed Transformer further extracts local and network-level information. The method introduces a Causal Restraint module to alleviate the correlation between features related to regions of interest (ROI) and irrelative features. Representative features are selected for the model through morphological and causal prior information, ultimately improving segmentation performance. [Results] On the public Synapse dataset, the Dice Similarity Coefficient and Hausdorff Distance achieved mean values of 80.01% and 19.39mm, respectively. On the public ACDC dataset, the mean DSC reached 90.95%, outperforming other comparative methods. [Conclusions] Experiment results demonstrate that the proposed method effectively enhances multi-organ segmentation performance on CT and MRI images and validates the feasibility of the causal restraint module across different models.

Key words: medical image segmentation, morphological constraint, Transformer, causal restraint