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

Late gadolinium enhancement cardiovascular magnetic resonance (LGE-CMR) permits non-invasive quantification of myocardial fibrosis; however, automated scar segmentation remains challenging due to limited expert annotations and reduced image quality caused by acquisition noise and artefacts. We investigate two related questions: (i) whether inversion of a stochastic Gaussian diffusion process can reproduce the texture characteristics of low-signal-to-noise-ratio (SNR) LGE imaging, and (ii) whether the resulting synthetic data can improve automated fibrosis segmentation in annotation-limited settings. To this end, we introduce a mask-conditioned denoising diffusion probabilistic model (DDPM) that synthesises high-fidelity 2D short-axis LGE-CMR slices from three-class label maps (background, myocardium, scar), and we employ these synthetic images for training-set augmentation. The impact of augmentation was assessed using the nnU-Net v2 segmentation framework and benchmarked against exemplar-guided image synthesis with CoCosNet-v2 under identical data partitioning. On a held-out test set trained with 100 real cases, inclusion of 300 diffusion-generated cases increased the scar Dice coefficient from 0.173 to 0.271 (+56.7%), and the scar recall from 0.173 to 0.363, demonstrating enhanced sensitivity to fibrotic lesions. For comparable training budgets, diffusion-based augmentation consistently outperformed GAN-based augmentation, although performance improvements were non-monotonic with respect to the real-to-synthetic data ratio and attenuated as the size of the real dataset increased. A four-axis noise-fidelity analysis (spectral content, signal-dependent variance, short-range spatial correlation, distributional shape) further shows that the DDPM reproduces scanner-specific noise statistics substantially more faithfully than the GAN baseline, providing a mechanistic account for the augmentation gap.