Multi-Label Conditioned Diffusion for Cardiac MR Image Augmentation and Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate cardiac MRI segmentation is vital for diagnosing heart diseases and planning treatments.
• High segmentation accuracy requires large, annotated datasets, which are difficult and expensive to acquire.
• Deep learning models for segmentation heavily depend on the quantity and quality of training data.

Purpose of the Study:

• To develop a novel data augmentation framework for cardiac MRI using a condition-guided diffusion generative model.
• To address the challenge of limited annotated cardiac MRI datasets.
• To improve the performance of cardiac segmentation tasks through enhanced datasets.

Main Methods:

• A two-stage generative data augmentation framework was proposed.
• Stage 1: A Label Diffusion Module generated realistic multi-category spatial masks based on anatomical priors.
• Stage 2: Cardiac MRI images were generated conditioned on these masks using a spatially-adaptive normalization (SPADE) module for structural accuracy.

Main Results:

• The generative augmentation framework significantly increased dataset sample numbers.
• Cardiac segmentation accuracy improved by 5% to 10% in Dice Similarity Coefficient (DSC) compared to traditional methods.
• A strong correlation was observed between the quality of generated images and augmentation effectiveness.

Conclusions:

• The proposed framework offers a robust solution for data scarcity in cardiac image analysis.
• Generative data augmentation using condition-guided diffusion models can substantially enhance downstream segmentation tasks.
• This approach has direct benefits for clinical applications in cardiac disease diagnosis and management.