DCL: Dynamic Causal Learning for Cross-Modality Cardiac Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate cardiac image segmentation is crucial for diagnosing and treating heart disease.
• Existing methods struggle with spatial-temporal confounding, hindering cross-modality knowledge transfer due to intertwined anatomy and modality features.

Purpose of the Study:

• To propose a novel Dynamic Causal Learning (DCL) method to overcome spatial-temporal confounding in cross-modality cardiac image segmentation.
• To improve the transferability of learned features across different cardiac imaging modalities (MR, CT, US).

Main Methods:

• Developed Dynamic Causal Learning (DCL) using multi-dimensional causal intervention to address spatial-temporal confounding.
• Integrated historical optimal interventions for knowledge transfer across temporal contexts.
• Employed a diffusion mechanism to ensure causal invariance of anatomical features.

Main Results:

• The DCL method achieved a mean Dice score of 0.951 on cross-modality cardiac images.
• DCL significantly outperformed existing advanced segmentation methods.
• Demonstrated effectiveness across multiple imaging modalities including MR, CT, and US.

Conclusions:

• Dynamic Causal Learning (DCL) effectively solves spatial-temporal confounding in cross-modality cardiac image segmentation.
• The proposed method enhances model performance and knowledge transfer across diverse cardiac imaging modalities.
• DCL offers a promising solution for improving cardiac image analysis and clinical decision-making.