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

Alexandra G. Roberts, Ha M. Luu, Mert Şişman, Alexey V. Dimov, Ceren Tozlu, Ilhami Kovanlikaya, Susan A. Gauthier, Thanh D. Nguyen, Yi Wang · 2025

Read Synthetic Generation and Latent Projection Denoising of Rim Lesions in Multiple Sclerosis by Alexandra G. Roberts, Ha M. Luu, Mert Şişman, Alexey V. Dimov, Ceren Tozlu, Ilhami Kovanlikaya, Susan A. Gauthier, Thanh D. Nguyen, Yi Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-23353.txt.

Introduction

Alexandra Roberts and co-authors tackle, in a 2025 paper, a data problem in multiple sclerosis imaging: the most informative lesions are also the rarest. Special MRI maps can reveal rim lesions, ringed with iron left by immune attack. These paramagnetic rim lesions are an emerging marker for prognosis and diagnosis. But they are scarce next to ordinary lesions, which starves classifiers.

The team generates synthetic versions of the susceptibility maps that carry rim lesions. Adding the synthetic cases to training improved classifier performance, they report. They also extend a denoising method based on latent projection to multiple channels. Cleaner maps and balanced data attack the same problem from both sides. Better rim lesion detection would let doctors track this aggressive lesion type over time. The work shows synthetic data earning its place in a clinical imaging pipeline.

Catalog record

CREATORAlexandra G. Roberts, Ha M. Luu, Mert Şişman, Alexey V. Dimov, Ceren Tozlu, Ilhami Kovanlikaya, Susan A. Gauthier, Thanh D. Nguyen, Yi Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-23353

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Synthetic Generation and Latent Projection Denoising of Rim Lesions in Multiple Sclerosis Alexandra G. Roberts, Ha M. Luu, Mert Şişman, Alexey V. Dimov, Ceren Tozlu, Ilhami Kovanlikaya, Susan A. Gauthier, Thanh D. Nguyen, Yi Wang 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. 1 Introduction Figure 1 : Example of MS patient with lesions depicted on qualitative \mathrm{T_{2}FLAIR} and whole brain QSM. Lesions appearing on both contrasts are indicated by arrows. 1.1 Background Multiple sclerosis (MS) is a debilitating neurodegenerative disease [ 40 ] with a rising global presence [ 60 ] . MS is the most common demyelinating disorder [ 62 ] and there is currently no cure [ 15 ] , though a variety of disease-modifying therapies exist to slow symptom progression and improve quality of life [ 48 ] . The efficacy of these therapies is measured by the appearance of gadolinium-enhancing lesions in brain white matter as identified on magnetic resonance imaging (MRI) [ 19 ] . The importance of MRI in both MS diagnosis and prognosis [ 30 ] motivates the development of automatic lesion detection and segmentation algorithms to address this growing clinical need [ 10 ] . 1.2 Imaging techniques Much progress on the understanding of MS has been gained through the qualitative \mathrm{T_{1}w} and \mathrm{T_{2}FLAIR} MRI contrasts [ 13 ] . However, longitudinal studies increasingly rely on quantitative susceptibility maps (QSM) as robust biomarkers in MS disease progression [ 68 , 14 , 7 , 58 ] . Of critical importance is the visualization of paramagnetic rim lesions linked to symptom severity [ 55 ] . Due to the iron depositions surrounding the rim of these lesions [ 21 ] , susceptibility contrasts like QSM are required to differentiate between lesion subtypes [ 22 ] . Co-registered \mathrm{T_{2}FLAIR} and whole brain QSM [ 46 ] are shown in an example MS patient in Figure 1 , with lesions indicated by the white arrows. Iron is involved in a variety of neuroinflammatory diseases and deposition often increases with inflammatory response [ 61 ] . Activated immune cell microglia at the edge of the lesion are the primary source of iron, generating contrast between the paramagnetic rim and diamagnetic lesion core [ 25 ] . As QSM directly measures small changes in the applied magnetic field arising from paramagnetic or diamagnetic tissue content, it is a reliable method to detect rim lesions [ 45 ] . Figure 2 illustrates the need for QSM to enable rim identification as compared to \mathrm{T_{2}FLAIR} . 1.3 Lesion class imbalance Though paramagnetic rim lesions differentiate between MS and other neurodegenerative disorders with high specificity [ 38 ] , only 10\% of all MS lesions are estimated to be rim lesions [ 43 ] , introducing a class imbalance problem for detection and segmentation algorithms. Despite consensus [ 4 ] on rim lesion characteristics, ambiguous cases remain when expert readers disagree on these criteria. It is desirable to “denoise” such lesions in order to make use of all possible rim lesion data. Figure 2 : Example of rim (top) and non-rim (bottom) lesion visualization on \mathrm{T_{2}FLAIR} (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. 1.4 Contributions We present a novel sample denoising approach based on generated synthetic quantitative susceptibility maps of paramagnetic rim lesions. We exploit the projection capability [ 8 ] in our trained generative network to ”denoise” ambiguous samples by recovering unambiguous synthesized samples. We evaluate the effect of training with said data in the rim classification problem and demonstrate improvement in trained detectors when denoised synthetic data is included. This approach enables us to train with noisy or ambiguous labels by recovering their synthetic analog from the trained generator network. To our knowledge, our proposed method is the first to allow ambiguous or contested rim lesion labels to augment the rare, unambiguous rim lesion class during rim detection. Finally, we provide a multi-contrast extension to enable generation of susceptibility maps, \mathrm{T_{2}FLAIR} images, and probabilistic rim lesion segmentations, broadening the utility of our framework. 2 Related Works 2.1 Synthetic MRI Synthetic MRI data ranges from classical representations generated via signal models [ 5 , 2 ] to more recent deep learning approaches aiming to estimate mappings over synthetic data that are applicable to in\ vivo data [ 42 , 17 ] and across various contrasts and resolutions [ 24 ] . Given pretrained models and approximate source and target distributions [ 63 ] , transfer learning approaches can address the challenges of gathering sufficient MRI data [ 33 , 57 , 39 ] . Another subset of solutions focus on generating synthetic data for small datasets [ 53 ] and label scarcity arising from new [ 59 ] or lethal [ 1 ] pathologies. Rim segmentation efforts have addressed the class imbalance problem by oversampling the minority class in the latent space [ 66 , 9 ] , included in our comparison. Synthetic lesions in multiple sclerosis have been generated [ 49 ] using a variational autoencoder on qualitative \mathrm{T_{2}FLAIR} . Like our work, this model aims to generate synthetic MS lesions, differing in that the whole brain learned mapping is between healthy cases and MS patients on qualitative \mathrm{T_{2}FLAIR} . We train a generative adversarial network [ 16 ] (GAN) with the goal of learning the mapping from random noise initializing the generator model to synthetic, quantitative, paramagnetic rim lesions. We do this using susceptibility images as rim lesions are differentiable from non-rim lesions only on QSM. 2.2 Latent projection denoising From our choice in architecture arises an opportunity to recover ambiguous or “noisy” rim lesions by projecting their latent vectors into the latent space of a pretrained GAN synthesizing only unambiguous rim lesions, closely related to GAN inversion [ 64 ] . Namely, these ambiguous cases are lesions where expert raters disagree on the label.

Read the complete text