English

SLIM-Diff: Shared Latent Image-Mask Diffusion with Lp loss for Data-Scarce Epilepsy FLAIR MRI

Computer Vision and Pattern Recognition 2026-02-04 v1 Artificial Intelligence

Abstract

Focal cortical dysplasia (FCD) lesions in epilepsy FLAIR MRI are subtle and scarce, making joint image--mask generative modeling prone to instability and memorization. We propose SLIM-Diff, a compact joint diffusion model whose main contributions are (i) a single shared-bottleneck U-Net that enforces tight coupling between anatomy and lesion geometry from a 2-channel image+mask representation, and (ii) loss-geometry tuning via a tunable LpL_p objective. As an internal baseline, we include the canonical DDPM-style objective (ϵ\epsilon-prediction with L2L_2 loss) and isolate the effect of prediction parameterization and LpL_p geometry under a matched setup. Experiments show that x0x_0-prediction is consistently the strongest choice for joint synthesis, and that fractional sub-quadratic penalties (L1.5L_{1.5}) improve image fidelity while L2L_2 better preserves lesion mask morphology. Our code and model weights are available in https://github.com/MarioPasc/slim-diff

Keywords

Cite

@article{arxiv.2602.03372,
  title  = {SLIM-Diff: Shared Latent Image-Mask Diffusion with Lp loss for Data-Scarce Epilepsy FLAIR MRI},
  author = {Mario Pascual-González and Ariadna Jiménez-Partinen and R. M. Luque-Baena and Fátima Nagib-Raya and Ezequiel López-Rubio},
  journal= {arXiv preprint arXiv:2602.03372},
  year   = {2026}
}

Comments

6 pages, 2 figures, 1 table, conference paper

R2 v1 2026-07-01T09:33:54.833Z