Modelling Stochastic Sensor Noise via Mask-Conditioned Diffusion for Data Augmentation in Low-SNR LGE-CMR

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiovascular Disease Research

Background:

• Late gadolinium enhancement cardiovascular magnetic resonance (LGE-CMR) is crucial for quantifying myocardial fibrosis.
• Automated scar segmentation in LGE-CMR is hindered by limited expert annotations and poor image quality (low signal-to-noise ratio, artifacts).
• Existing methods struggle with the inherent noise and texture characteristics of low-SNR LGE imaging.

Purpose of the Study:

• To investigate if a diffusion model can replicate LGE-CMR texture characteristics.
• To determine if synthetic LGE-CMR data generated by diffusion models can improve automated fibrosis segmentation in data-limited scenarios.
• To compare the effectiveness of diffusion-based augmentation against Generative Adversarial Network (GAN)-based augmentation.

Main Methods:

• Introduced a mask-conditioned denoising diffusion probabilistic model (DDPM) to synthesize 2D LGE-CMR slices from label maps.
• Employed synthetic images for training-set augmentation in the nnU-Net v2 segmentation framework.
• Benchmarked against exemplar-guided image synthesis (CoCosNet-v2) and evaluated noise-fidelity characteristics.

Main Results:

• Augmenting a real dataset with 300 diffusion-generated cases significantly improved scar Dice coefficient (+56.7%) and recall on a held-out test set.
• Diffusion-based augmentation consistently outperformed GAN-based augmentation for comparable training budgets.
• The DDPM demonstrated superior fidelity in reproducing scanner-specific noise statistics compared to the GAN baseline.

Conclusions:

• Diffusion models can effectively generate realistic LGE-CMR images, addressing challenges of limited annotations and low image quality.
• Synthetic data from DDPMs substantially enhances automated fibrosis segmentation performance, particularly in annotation-limited settings.
• The DDPM offers a more mechanistically sound approach to data augmentation for LGE-CMR analysis compared to GANs.