Lesion-DDPM: Lesion-Enhanced 3D Diffusion for MS MRI Synthesis

arXiv:2606.15457 · cs.CV, cs.LG · Submitted 2026-06-13 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Lesion-DDPM: Lesion-Enhanced 3D Diffusion for MS MRI Synthesis".

Tom: The gist:

Jane: First, who's behind it and why it matters.

Title and authors: Tom: Let’s start with who is behind this work. We’re talking about Lesion-DDPM: Lesion-Enhanced three dee Diffusion for MS MRI Synthesis by Zhang, Jung, Anik, Xiao, Janarthanan, Chuluunbaatar, Ho, Lee and Ryoo <ref:2606.15457#pg1,Lesion-DDPM: Lesion-Enhanced 3D Diffusion for MS MRI Synthesis>.

Jane: It’s a team from Northeastern University and Fairleigh Dickinson University collaborating with folks from the University of Colorado at Colorado Springs and National Taiwan University. It shows a nice mix of expertise here.

Lu: The authors are clearly focused on bridging the gap between standard diffusion models, which are great for general image generation, and the specific needs of medical imaging where structure matters.

Meng: What’s interesting is how they frame it—they aren't just making pretty pictures; they’re doing this for a clinical use case with a clear need: better lesion visualization.

Lalam: This paper sets up a framework that builds on existing work in GANs and diffusion models, but it steers the process to prioritize those lesions during the generation steps.

The paper's summary: Tom: Now for what Lesion-DDPM actually proposes. They frame it as a conditional Gaussian diffusion model designed for three dee FLAIR MRI synthesis <ref:2606.15457#pg1>. It uses a forward process where noise is added gradually according to a cosine variance schedule, creating those noisy volumes.

Jane: And at every step of that diffusion process, they use this three dee U-Net denoiser to predict the noise field based on the current noisy image and some conditions <ref:2606.15457#pg1>.

Lu: The key part here is how they condition that denoiser. They use a three dee U-Net architecture with skip connections, which helps it recover those fine anatomical details while it’s denoising <ref:2606.15457#pg1>.

Meng: But there’s more than just the basic architecture; they introduce multi-level gated condition injection, which means the anatomical mask is injected at several points throughout the network.

Lalam: This injection happens by resizing and encoding that mask at each encoder stage and decoder stage into a scale-matched feature map, then fusing it with the main stream using a channel-wise concatenation followed by a point-wise three dee convolution and a learnable sigmoid gate <ref:2606.15457#pg1>.

The paper's improvements: Tom: So what are the actual improvements they claim? It’s two main things that make this method stand out in terms of performance. First, they use a lesion-weighted L1 loss.

Jane: That means the objective function is modified so it increases the gradient emphasis on lesion voxels while maintaining overall image fidelity in non-lesion regions.

Lu: By up-weighting those specific voxels relative to tissue and background, they are directly addressing that issue where generative models tend to oversmooth subtle lesions or introduce artifacts.

Meng: That weighting mechanism is designed to increase the gradient emphasis on lesion voxels while maintaining overall image fidelity in non-lesion regions. It’s a direct attempt to solve the lesion–background imbalance problem.

Lalam: They also use multi-level gated condition injection, which helps reinforce lesion context across multiple feature scales and improves conditioning stability. That seems like it gives the system more control over what it prioritizes during synthesis.

Conclusion: Tom: So to wrap up, Lesion-DDPM is this three dee conditional diffusion framework that uses both a weighted loss and multi-level condition injection to synthesize lesion-aware FLAIR MRI volumes <ref:2606.15457#pg1>.

Jane: The results show it achieved the lowest lesion-region reconstruction error among the methods they compared against. It also showed strong performance in downstream tasks like three dee U-Net lesion segmentation, reaching a Dice score of zero point six one six when trained only on its own generated scans and evaluated on real MRIs.

Lu: The synthetic images generated by Lesion-DDPM were often perceived as real by neurologists, suggesting high visual realism and diagnostic plausibility in their output.

Meng: Quantitatively, it yielded the lowest lesion-wise mean squared error of zero point zero four eight and a mean absolute error of zero point one eight one, along with the highest MS-SSIM score of zero point seven nine eight compared to other baselines they tested on the MSLesSeg dataset.

Lalam: This work really highlights the potential for this model for data augmentation and lesion-preserving MRI synthesis under limited data conditions, which is a big deal in medical imaging right now.

Tom: It’s a solid paper that shows how targeted loss functions and detailed conditioning can specifically preserve subtle, critical structures in complex three dee data like brain MRIs <ref:2606.15457#pg1>.

Jane: Definitely something to keep an eye on as we look at how these conditional diffusion techniques evolve for other medical image modalities next.

Northeastern University, Vancouver, BC · Fairleigh Dickinson University, Vancouver, BC · University of Colorado at Colorado Springs, Colorado Springs, CO · National Taiwan University, Taipei, Taiwan · University of British Columbia, Vancouver, BC · British Columbia Institute of Technology

cs.CV, cs.LG

Submitted: 2026-06-13

Updated: 2026-10-08

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: The gist: Lesion-DDPM proposes a 3D conditional diffusion framework for lesion-aware FLAIR synthesis that incorporates multi-level anatomical mask injection and a lesion-weighted reconstruction loss

Key concepts

3D Conditional Diffusion Framework: This is the core model type used for image generation. It works by gradually adding noise to an image (forward process) and then training a neural network (a 3D U-Net denoiser) to reverse that noise, guided by both the noisy input and a condition like an anatomical mask, ensuring the final output matches the desired structure.
Lesion-Weighted Reconstruction Loss: This is a specific loss function used during training. It adjusts how much the model focuses on reconstructing different parts of the image. By weighting lesion voxels more heavily than background tissue, it forces the model to pay close attention to small MS lesions, preventing them from being oversmoothed or lost during synthesis.
Multi-Level Gated Condition Injection: This technique involves injecting anatomical mask information into the diffusion process at several points—the encoder stages, bottleneck, and decoder stages. This injection uses a learned sigmoid gate to control how much the structural guidance influences the denoising step at each level, providing more precise spatial control over lesion placement.
Lesion-Aware FLAIR Synthesis: This is the goal of the research: creating synthetic 3D FLAIR MRI scans specifically tailored for MS. The synthesis aims to accurately reproduce both the general brain anatomy and, crucially, the small, sparse hyperintensities characteristic of MS lesions, addressing the challenge of lesion-background imbalance.

Terminology

Generative Model — 1× in this paper · explained in 2 episodes
Generative Models — 1× in this paper · explained in 3 episodes
Reconstruction Error — 1× in this paper · explained in 2 episodes

Summary

The gist: Lesion-DDPM proposes a 3D conditional diffusion framework for lesion-aware FLAIR synthesis that incorporates multi-level anatomical mask injection and a lesion-weighted reconstruction loss to emphasize lesion voxels while maintaining global brain structure.

Introduction and Motivation

Clinical brain MRI for multiple sclerosis (MS) is typically performed using multisequence protocols, but contemporary guidelines emphasize sagittal 3D FLAIR (fluid-attenuated inversion recovery) as a core sequence for diagnosis and monitoring due to its high sensitivity to supratentorial lesions and new lesion activity By suppressing cerebrospinal fluid (CSF) signal, FLAIR improves the conspicuity of periventricular and juxtacortical hyperintensities—two hallmark lesion locations in MS—and can enhance visual detection compared to conventional T2-weighted imaging in these regions Moreover, 3D FLAIR generally yields higher lesion counts and better contrast-to-noise ratios than 2D FLAIR, with multiple studies reporting substantially higher lesion detection across both supratentorial and infratentorial compartments Motivated by its clinical role, we target 3D FLAIR as the synthesis objective and aim to generate high-fidelity synthetic volumes that preserve small MS lesions MS lesions are small, sparse, and morphologically irregular; in voxel space they occupy only a tiny fraction of brain volume, leading to severe lesion–background imbalance These characteristics make lesion visibility difficult to preserve: generative models may oversmooth subtle lesions or introduce artifacts

Model Architecture

Lesion-DDPM is formulated as a conditional Gaussian diffusion model for 3D FLAIR MRI synthesis The forward process gradually adds Gaussian noise according to a cosine variance schedule, producing a sequence of noisy volumes At each diffusion step t, the reverse transition pθ(xt−1 xt, c) is parameterized by a 3D U-Net denoiser that predicts the noise field ϵθ(xt, t, c) The denoiser adopts a 3D U-Net architecture with skip connections between symmetric encoder and decoder stages A sinusoidal timestep embedding is fed into every residual block, conditioning predictions on the current noise level, while the anatomical mask c provides semantic guidance on brain and lesion locations Beyond optional input-level concatenation of xt and c, Lesion-DDPM introduces multi-level gated condition injection At each encoder stage, the bottleneck, and each decoder stage, the mask is resized and encoded into a scale-matched feature map This conditional feature is then fused into the main feature stream by channel-wise concatenation followed by a point-wise 3D convolution and a learnable sigmoid gate that scales the residual update before it is added back to the original feature map Training uses the standard diffusion noise-prediction objective with a lesion-weighted L1 loss that up-weights lesion voxels relative to tissue and background

Experimental Setup and Comparison

The experiments utilized the MSLesSeg dataset We compared Lesion-DDPM against three representative models (3D Pix2Pix, 3D DiscoGAN, Med-DDPM) where each was trained with the same data splits and preprocessing We fixed the total training steps for all methods to match the training budget and ensure a fair comparison Lesion-DDPM adopted a two-stage schedule (15,000 steps at 1× 10−5 followed by 10,000 steps at 1 × 10−6) We further performed ablations on multi-level gated condition injection and the lesion-weighted loss

Results and Evaluation

Lesion-DDPM achieved the lowest lesion-region reconstruction error among all methods In a downstream 3D U-Net lesion segmentation task, a model trained only on LesionDDPM–generated scans and evaluated on real MRIs reached a Dice score of 0.616 compared with 0.569 for the best competing synthetic dataset Lesion-DDPM again demonstrated clear superiority, yielding the lowest lesion-wise MSE (0.048) and MAE (0.181), along with the highest MS-SSIM (0.798) In a hybrid configuration, Lesion-DDPM achieved the highest Dice (0.685) and IoU (0.534), while 3D DiscoGAN attained the highest recall (0.781)

Conclusion

In this study, we presented Lesion-DDPM, a 3D conditional diffusion framework specifically designed for lesion-aware FLAIR MRI synthesis in multiple sclerosis By incorporating multi-level conditional injection and a lesion-weighted L1 objective, our method effectively preserved small lesion structures while maintaining global anatomical consistency Expert assessment demonstrated that synthetic images generated by Lesion-DDPM were often perceived as real by neurologists, indicating high visual realism and diagnostic plausibility Quantitative evaluations further showed that LesionDDPM outperformed existing GAN-based and diffusion-based baselines across MSE, MAE, and MS-SSIM metrics These findings highlight the model’s potential for data augmentation and lesion-preserving MRI synthesis under limited data conditions Future extensions may explore multi-modal MRI synthesis, such as incorporating T2-weighted or crosssequence generation to enhance anatomical consistency and cross-modality generalization

References

[1] M. P. Wattjes, O. Ciccarelli, D. S. Reich, B. Banwell, N. de Stefano, C. Enzinger, F. Fazekas, M. Filippi, J.- Frederiksen, C. Gasperini et al., “2021 magnims–cmsc–naims consensus recommendations on the use of mri in patients with multiple sclerosis,” The Lancet Neurology, vol. 20, no. 8, pp. 653–670, 2021

[2] R. Bakshi, S. Ariyaratana, R. H. Benedict, and L. Jacobs, “Fluidattenuated inversion recovery magnetic resonance imaging detects cortical and juxtacortical multiple sclerosis lesions,” Archives of neurology, vol. 58, no. 5, pp. 742–748, 2001 <ref:2606.15457#pg2>

[3] A. Bink, M. Schmitt, J. Gaa, J. P. Mugler 3rd, H. Lanfermann, and F. E. Zanella, “Detection of lesions in multiple sclerosis by 2d flair and single-slab 3d flair sequences at 3.0 t: initial results,” European radiology, vol. 16, no. 5, pp.

Improvements for AI systems

The Lesion-DDPM framework can synthesize 3D FLAIR MRI volumes that preserve small MS lesions by employing a lesion-weighted L1 reconstruction loss that up-weights lesion voxels relative to tissue and background. This allows the system to generate synthetic data with superior lesion fidelity compared to baselines.
The model's multi-level gated condition injection enables it to reinforce lesion context across multiple feature scales and improving conditioning stability, allowing the AI system to maintain whole-brain realism while explicitly prioritizing subtle lesion fidelity.
Lesion-DDPM can be used for downstream 3D U-Net segmentation where a model trained only on its generated scans reached a Dice score of 0.616 and further increased to 0.685 when real images were added to the training set, confirming its utility in enhancing segmentation robustness.
The system can generate synthetic cohorts with structural consistency close to real data, achieving a set-level MS-SSIM of 0.660, suggesting it produces synthetic images whose overall anatomical structures and lesion configurations are closer to those of the real dataset.

Abstract

3D FLAIR MRI is widely recommended as one of the standard MRI sequences for brain imaging in multiple sclerosis (MS), but publicly available MS datasets remain relatively small and vary across scanners, acquisition protocols, and lesion patterns. This scarcity and variability hinder the development of robust neuroimaging machine learning models and are particularly challenging for generative models that aim to synthesize images while preserving small, sparse lesions. We propose Lesion-DDPM, a 3D conditional diffusion framework for lesion-aware FLAIR synthesis that incorporates multi-level anatomical mask injection together with a lesion-weighted reconstruction loss to emphasize lesion voxels while maintaining global brain structure. Using a curated subset of the MSLesSeg dataset, we compare Lesion-DDPM with representative state-of-the-art GAN- and diffusion-based models, assessing both image-generation metrics and downstream 3D U-Net segmentation. In our experiments, Lesion-DDPM achieved the lowest lesion-region reconstruction error among all methods. In a downstream 3D U-Net lesion segmentation task, a model trained only on Lesion-DDPM-generated scans and evaluated on real MRIs reached a Dice score of 0.616 compared with 0.569 for the best competing synthetic dataset. When Lesion-DDPM images were added to the real training set, the Dice score further increased to 0.685.