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

Technical Summary: Lesion-DDPM

Problem Statement
Multiple Sclerosis (MS) diagnosis and monitoring rely heavily on 3D FLAIR MRI sequences due to their high sensitivity to supratentorial lesions. However, the development of robust neuroimaging machine learning models is hindered by the scarcity and heterogeneity of publicly available MS datasets, which vary significantly across scanners and acquisition protocols. This scarcity is particularly challenging for generative models tasked with synthesizing 3D volumes. MS lesions are characteristically small, sparse, and morphologically irregular, occupying a tiny fraction of the brain volume. This severe lesion-background imbalance causes existing generative models (such as GANs and standard diffusion models) to either oversmooth subtle lesions, introduce artifacts, or fail to preserve lesion visibility while maintaining global brain structure. Current approaches often struggle to balance whole-brain anatomical realism with the explicit preservation of these tiny pathological regions.

Methodology
The authors propose Lesion-DDPM, a 3D conditional diffusion framework designed specifically for lesion-aware FLAIR synthesis. The model is formulated as a conditional Gaussian diffusion process where the target is a preprocessed, skull-stripped FLAIR volume, and the condition is an anatomical mask encoding brain and lesion regions.

The architecture and training strategy incorporate two primary technical innovations:

1. Multi-Level Gated Condition Injection: Unlike standard conditional diffusion models that may only concatenate masks at the input, Lesion-DDPM injects anatomical mask information at multiple stages: the encoder, the bottleneck, and the decoder. At each stage, the mask is resized to match the feature resolution, encoded into a feature map, and fused with the main stream via channel-wise concatenation and a pointwise 3D convolution. Crucially, this fusion is modulated by a learnable sigmoid gate, allowing the network to control the strength of lesion information injection at different depths. This design ensures that lesion cues inform both high-level global representations and low-level local details without over-conditioning.
2. Lesion-Weighted Reconstruction Loss: To address the severe class imbalance between lesion and background voxels, the authors replace the standard noise-prediction objective with a lesion-weighted L1 loss. This objective assigns a substantially higher weight to errors occurring within lesion voxels compared to tissue and background voxels. This forces the model to prioritize gradient updates for lesion regions, preventing the iterative denoising process from smoothing out small, sparse lesions.

Key Contributions

• Lesion-Aware Framework: The paper introduces a 3D conditional diffusion model that explicitly prioritizes the fidelity of small, sparse MS lesions while maintaining global anatomical consistency.
• Architectural Novelty: The implementation of multi-level gated condition injection allows for the reinforcement of lesion context across multiple feature scales, improving conditioning stability.
• Optimization Strategy: The integration of a lesion-weighted L1 reconstruction loss specifically targets the preservation of lesion voxels, mitigating the common issue of lesion oversmoothing in generative models.

Results
The model was evaluated using a curated subset of the MSLesSeg dataset (80 training pairs, 20 testing pairs) and compared against representative state-of-the-art baselines: 3D Pix2Pix, 3D DiscoGAN (GAN-based), and Med-DDPM (diffusion-based).

• Visual and Quantitative Fidelity: In expert assessments, synthetic images generated by Lesion-DDPM were frequently indistinguishable from real images by neurologists. Quantitatively, Lesion-DDPM achieved the lowest Mean Squared Error (MSE: 0.059) and Mean Absolute Error (MAE: 0.202) and the highest Multi-Scale Structural Similarity (MS-SSIM: 0.806) among all methods. Notably, in lesion-specific regions, it achieved the lowest reconstruction error (MSE: 0.048), significantly outperforming baselines.
• Downstream Segmentation Performance: A 3D U-Net trained exclusively on Lesion-DDPM-generated synthetic data achieved a Dice score of 0.616 on real MRI test data, outperforming the best competing synthetic dataset (Med-DDPM at 0.569).
• Hybrid Training: When Lesion-DDPM synthetic images were added to a real training set (hybrid training), the downstream segmentation Dice score increased to 0.685, surpassing the performance of models trained on real data alone (0.670) and other hybrid configurations.

Significance and Claims
The paper claims that Lesion-DDPM successfully addresses the challenge of generating high-fidelity synthetic 3D FLAIR volumes that preserve subtle MS lesions, a task where previous generative models have struggled due to lesion-background imbalance. The authors assert that their method produces visually convincing images that are diagnostically plausible, as evidenced by expert assessment. Furthermore, the results demonstrate that Lesion-DDPM-generated data is not only visually realistic but also functionally effective for training downstream segmentation models, offering a viable solution for data augmentation in limited-data scenarios. The study concludes that the combination of multi-level gated injection and lesion-weighted loss is critical for achieving this balance between global structure and local lesion fidelity. Future work suggested by the authors includes exploring multi-modal synthesis and controllable lesion progression simulation.