Synthetic Generation and Latent Projection Denoising of Rim Lesions in Multiple Sclerosis

TL;DR

This work tackles the challenge of detecting paramagnetic rim lesions (PRLs) in multiple sclerosis by leveraging quantitative susceptibility maps (QSM) to distinguish rim from non-rim lesions. It introduces a pipeline that synthesizes PRLs with StyleGAN2-ADA, employs latent-space projection denoising to augment the minority class with denoised ambiguous cases, and extends the framework to multi-contrast generation (susceptibility, T2FLAIR, and probabilistic segmentation). Radiologist-guided assessment validates the realism of synthetic rims and the denoising approach improves classifier performance and interpretability, achieving closer alignment to unseen rim distributions. The approach offers a practical path to mitigate class imbalance in PRL detection, enhances clinical relevance through interpretable maps, and provides resources for broader extension to multi-contrast MRI analyses and segmentation tasks.

Abstract

Quantitative susceptibility maps from magnetic resonance images can provide both prognostic and diagnostic information in multiple sclerosis, a neurodegenerative disease characterized by the formation of lesions in white matter brain tissue. In particular, susceptibility maps provide adequate contrast to distinguish between "rim" lesions, surrounded by deposited paramagnetic iron, and "non-rim" lesion types. These paramagnetic rim lesions (PRLs) are an emerging biomarker in multiple sclerosis. Much effort has been devoted to both detection and segmentation of such lesions to monitor longitudinal change. As paramagnetic rim lesions are rare, addressing this problem requires confronting the class imbalance between rim and non-rim lesions. We produce synthetic quantitative susceptibility maps of paramagnetic rim lesions and show that inclusion of such synthetic data improves classifier performance and provide a multi-channel extension to generate accompanying contrasts and probabilistic segmentation maps. We exploit the projection capability of our trained generative network to demonstrate a novel denoising approach that allows us to train on ambiguous rim cases and substantially increase the minority class. We show that both synthetic lesion synthesis and our proposed rim lesion label denoising method best approximate the unseen rim lesion distribution and improve detection in a clinically interpretable manner. We release our code and generated data at https://github.com/agr78/PRLx-GAN upon publication.

Figures (10)

• Figure 1: Example of MS patient with lesions depicted on qualitative $\mathrm{T_2FLAIR}$ and whole brain QSM. Lesions appearing on both contrasts are indicated by arrows.
• Figure 2: Example of rim (top) and non-rim (bottom) lesion visualization on $\mathrm{T_2FLAIR}$ (left) and QSM (right). Note that rim lesions can only be differentiated from non-rim lesions on QSM. The rim lesion is composed of a hypointense core (black arrow) and hyperintense rim (white arrow), contrast to a hyperintense non-rim lesion.
• Figure 3: Schematic of the GAN with adaptive discriminator augmentation blocks used to generate synthetic rim lesion data from limited datasets. The architecture consists of generative and discriminative networks $g(\mathbf{z},\mathbf{\theta})$ and $d(.,\phi)$ and receives inputs real images $\mathbf{x}$ and latent variable $\mathbf{z} \sim \mathcal{N}(0,1)$. Adapted from Tker.
• Figure 4: Outline of denoising approach, beginning with generative model training on unambiguous rim lesion $\mathbf{x}$. After training, features of ambiguous rim lesions (with noisy labels) are extracted to latent variable $\mathbf{w}_{\tilde{\mathbf{x}}}$ and projected onto the unambiguous latent space $\mathcal{W}$. Finally, the real rim lesion training data is augmented with these "denoised" rim lesions and the classifier performance is seen to improve.
• Figure 5: A simplified representation (adapted from Tker1) of the mapping $m$ and synthesis $s$ modules used to enable denoising via the trained generative network $g$. Given an ambiguous rim $\tilde{\mathbf{x}}$ lesion, denoising occurs via latent projection $\mathcal{P}$ after feature extraction by pretrained network $f$. The denoised projection $\mathbf{w}^*$ results from minimizing the perceptual loss $L_2$ term ($L_P$) with added noise regularization (Equation \ref{['eq1']}). After training with the architecture in Figure \ref{['f3']}, the synthesis module $s$ decodes the denoised projection $\hat{\mathbf{x}}$ corresponding to the "closest" unambiguous rim lesion in the latent space $\mathcal{W}$, enabling the augmentation in Figure \ref{['f4']}.