POWDR: Pathology-preserving Outpainting with Wavelet Diffusion for 3D MRI

TL;DR

POWDR tackles the scarcity of pathology-rich 3D MRI data by introducing a pathology-preserving outpainting framework that conditions generation on real pathology in the wavelet domain. By integrating a conditioned wavelet diffusion model with random-connected mask training and a 3D ResUNet backbone, it achieves diverse, anatomically plausible syntheses that preserve lesions while expanding surrounding tissue. Quantitative metrics (FID, MS-SSIM, LPIPS) show realism, and downstream clinical assessments (nnU-Net segmentation and tissue-volume analysis) demonstrate practical benefits and anatomical plausibility. The approach is tissue-agnostic, demonstrated on brain and knee MRI, and holds promise as a scalable data-augmentation tool to improve robustness in medical imaging pipelines.

Abstract

Medical imaging datasets often suffer from class imbalance and limited availability of pathology-rich cases, which constrains the performance of machine learning models for segmentation, classification, and vision-language tasks. To address this challenge, we propose POWDR, a pathology-preserving outpainting framework for 3D MRI based on a conditioned wavelet diffusion model. Unlike conventional augmentation or unconditional synthesis, POWDR retains real pathological regions while generating anatomically plausible surrounding tissue, enabling diversity without fabricating lesions. Our approach leverages wavelet-domain conditioning to enhance high-frequency detail and mitigate blurring common in latent diffusion models. We introduce a random connected mask training strategy to overcome conditioning-induced collapse and improve diversity outside the lesion. POWDR is evaluated on brain MRI using BraTS datasets and extended to knee MRI to demonstrate tissue-agnostic applicability. Quantitative metrics (FID, SSIM, LPIPS) confirm image realism, while diversity analysis shows significant improvement with random-mask training (cosine similarity reduced from 0.9947 to 0.9580; KL divergence increased from 0.00026 to 0.01494). Clinically relevant assessments reveal gains in tumor segmentation performance using nnU-Net, with Dice scores improving from 0.6992 to 0.7137 when adding 50 synthetic cases. Tissue volume analysis indicates no significant differences for CSF and GM compared to real images. These findings highlight POWDR as a practical solution for addressing data scarcity and class imbalance in medical imaging. The method is extensible to multiple anatomies and offers a controllable framework for generating diverse, pathology-preserving synthetic data to support robust model development.

Figures (6)

• Figure 1: Architecture of POWDR: a pathology-preserving outpainting framework using conditioned wavelet diffusion model for 3D MRI. The pipeline begins with wavelet decomposition and Gaussian noise addition during forward diffusion. In the conditional denoising step, the wavelet-transformed masked pathology is concatenated with the noisy input and processed by an attention-based ResUNet. The inverse wavelet transform recombines frequency components to reconstruct the synthesized image
• Figure 2: Qualitative example of brain MRI outpainting. Shown are A) the input condition (masked tumor), B) the synthetic image generated by POWDR, and C) the original image from which the tumor was extracted. Highlighted boxes indicate tumor regions preserved during synthesis. D) Synthetic image generated using a conditional Latent Diffusion outpainting model failed to converge.
• Figure 3: Sampling diversity under different training strategies. A–B) Mean, standard deviation, and example outputs from 20-case repeated sampling using A) masked tumor conditioning and B) random connected mask conditioning. C) Similarity metrics cosine similarity and KL divergence quantify diversity improvements.
• Figure 4: Impact of synthetic augmentation on tumor segmentation. Dice scores for nnU-Net segmentation under varying numbers of synthetic cases (10, 50, 100) compared to baseline.
• Figure 5: Comparison of brain tissue volumes between real and synthetic images. Volumes of CSF, gray matter, and white matter segmented using FSL FAST for 100 synthetic and 100 real cases.