Self-supervised Multimodal Transformer for Accurate Low-dose CT Liver Tumor Segmentation with Cross-domain Adaptation

Cardiovascular diseases, as one of the major causes of death worldwide, are in urgent need of high-precision and high-resolution techniques for the assessment of cardiac structure and function. Dynamic cardiac magnetic resonance imaging (dMRI), as a non-invasive imaging technique, has the unique advantage of capturing the dynamic changes in the full cycle of cardiac motion, however, it still faces the limitation of the trade-off between temporal resolution and image quality in practical applications. To address the above bottlenecks, this study proposes a multi-task diffusion modeling framework (CMDM) based on causal modeling, which combines a deep diffusion network with a causal inference method to collaborate on the tasks of super-resolution reconstruction and functional assessment of cardiac images in the same network. The model integrates a causal attention mechanism based on Granger causality theory through an encoder-diffuser core architecture and a dual-task-specific head design, which enhances the accurate recognition of cardiac function while preserving the details of image spatial structure. Extensive experiments on multiple clinical datasets, including ACDC and UK Biobank, show that CMDM significantly outperforms existing state-of-the-art methods in key metrics, such as PSNR, SSIM, and LVEF, and that embedding causal inference into multi-task learning provides better interpretability and performance for complex medical imaging scenarios.