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

TL;DR

This work tackles data scarcity in epilepsy-specific FLAIR MRI by introducing SLIM-Diff, a diffusion-based framework for joint image and lesion-mask synthesis. It uses a compact, shared-bottleneck U-Net to jointly model image and mask with a 2-channel input and applies a tunable $L_p$ loss across prediction targets, revealing that $x_0$-prediction generally yields best image quality while $L_2$ optimizes lesion-morphology. The study demonstrates that sub-quadratic losses like $L_{1.5}$ improve image fidelity, whereas $L_2$ better preserves mask geometry, all within a 26.9M-parameter model designed for data-scarce regimes. The approach is validated with distributional and morphology metrics (KID, LPIPS, MMD-MF) and rigorous non-parametric statistics, and opens avenues for data augmentation in clinical settings, albeit with limitations in volumetric consistency due to 2D slicing.

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 $L_p$ objective. As an internal baseline, we include the canonical DDPM-style objective ($ε$-prediction with $L_2$ loss) and isolate the effect of prediction parameterization and $L_p$ geometry under a matched setup. Experiments show that $x_0$-prediction is consistently the strongest choice for joint synthesis, and that fractional sub-quadratic penalties ($L_{1.5}$) improve image fidelity while $L_2$ better preserves lesion mask morphology. Our code and model weights are available in https://github.com/MarioPasc/slim-diff

Figures (2)

• Figure 1: Overview of SLIM-Diff. (A) Training strategy. (B) Joint image--mask synthesis architecture (details in Section \ref{['sec:shared-bottleneck']}) (C) Evaluation under axial-depth and pathology conditioning.
• Figure 2: Similarity and mask-quality metrics for SLIM-Diff across prediction targets ($\epsilon$, $v$, $x_0$) and $L_p$ settings. (A) Image realism is quantified with KID and LPIPS (lower is better), reported against a held-out real test set and contextualized with a real-vs-real baseline (two disjoint subsets of real data under the same protocol). (B) Lesion mask realism is quantified with MMD-MF (Maximum Mean Discrepancy on Morphological Features) and complemented with per-feature Wasserstein distances over nine standard shape descriptors; the best configuration is highlighted in the figure.