⚡ electrical engineering

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

POWDR is a pathology-preserving outpainting framework for 3D MRI that utilizes a conditioned wavelet diffusion model with random connected mask training to generate diverse, anatomically plausible synthetic data without fabricating lesions, thereby effectively addressing class imbalance and improving downstream segmentation performance.

Original authors: Fei Tan, Ashok Vardhan Addala, Bruno Astuto Arouche Nunes, Xucheng Zhu, Ravi Soni

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fei Tan, Ashok Vardhan Addala, Bruno Astuto Arouche Nunes, Xucheng Zhu, Ravi Soni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize a rare type of tumor in a brain scan. The problem is, you only have a few pictures of brains with that specific tumor, but you have thousands of pictures of healthy brains. It's like trying to teach someone to recognize a specific rare bird by showing them one photo of the bird and a million photos of pigeons. The robot gets confused and doesn't learn the rare bird well.

Usually, scientists try to fix this by making up fake pictures of the rare bird. But here's the catch: if you just ask an AI to "draw a brain with a tumor," it often makes up a tumor that doesn't look real, or it changes the real tumor you wanted to keep. It's like asking an artist to draw a bird, but they accidentally change the bird's beak or wings while trying to add the background.

Enter POWDR: The "Smart Outpainter"

The authors of this paper created a tool called POWDR (Pathology-preserving Outpainting with Wavelet Diffusion). Think of it as a very special "fill-in-the-blanks" artist for 3D medical scans.

Here is how it works, using simple analogies:

1. The "Do Not Touch" Zone (Pathology Preservation)

Imagine you have a photo of a brain with a tumor. You take a piece of tape and cover the tumor, but you leave the rest of the brain exposed. You tell the AI: "Keep the tumor exactly as it is. I don't want you to change it. But please, fill in the rest of the brain around it with new, realistic tissue."

This is called outpainting. Unlike other methods that might try to redraw the whole picture (and accidentally mess up the tumor), POWDR locks the real tumor in place and only generates the healthy tissue around it. This ensures the "rare bird" stays exactly the same while the "sky" around it changes.

2. The "High-Definition" Trick (Wavelet Diffusion)

Most AI image generators are like blurry photocopiers; they are good at the big shapes but lose the tiny, sharp details (like the texture of brain tissue).

POWDR uses a technique called Wavelet Diffusion. Imagine looking at a painting through a special pair of glasses that separates the image into "big, blurry shapes" and "tiny, sharp details."

The AI first learns to draw the big, blurry shapes.
Then, it uses the "glasses" to focus specifically on the tiny, sharp details (the high-frequency parts) to make sure the new tissue looks crisp and real, not fuzzy.

3. The "Random Mask" Game (Solving the Copy-Paste Problem)

The authors noticed a problem: if you always tell the AI to fill in the space around the exact same tumor, the AI gets lazy and just copies the same background every time. It's like asking a chef to make a sandwich with the same bread and ham every day; eventually, every sandwich looks identical.

To fix this, they introduced a Random Connected Mask strategy.

Instead of just covering the tumor, they sometimes cover random, connected chunks of the brain (like a puzzle piece) and ask the AI to fill those in.
This forces the AI to learn how to create many different types of healthy brain tissue, not just one version.
The Result: When they tested this, the AI stopped making identical copies. The "diversity" score went up, meaning the fake brains looked much more varied and realistic, just like real patients.

4. The Proof: Does it actually help?

The team tested POWDR on two things: Brain Scans (using a famous dataset called BraTS) and Knee Scans.

The Brain Test: They used POWDR to create 50 extra fake brain scans with tumors. They then used these fake scans to train a robot (an AI called nnU-Net) to find tumors.
- The Result: The robot got better at finding tumors! Its accuracy score (Dice score) went up from about 0.70 to 0.71. It's a small jump, but in the world of medical AI, that's a big win.
The Knee Test: They tried it on knee scans with torn cartilage. Even though knees look very different from brains, the AI worked just as well. This proves the method isn't just a "brain trick"; it works on different body parts too.
The "Realness" Check: They measured the volume of fluid and tissue in the fake brains. The amounts were almost identical to real human brains, proving the AI didn't make the brains too big or too small.

Summary

POWDR is a new way to create fake medical images that helps solve the problem of not having enough data. It does this by:

Locking the real disease in place so it doesn't get changed.
Generating new, realistic healthy tissue around it.
Using special math (wavelets) to keep the image sharp and detailed.
Forcing variety so the fake images don't all look the same.

The paper claims this is a practical way to make more training data for medical AI, helping them learn better without needing to scan more real patients. They tested it on brains and knees, and it worked well for both.

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

Problem Statement

The development of robust deep learning models for medical imaging is significantly hindered by two primary challenges: the scarcity of pathological data and pronounced class imbalance. Acquiring large-scale, diverse datasets is resource-intensive, and many pathological conditions occur at low prevalence, limiting the availability of representative samples. While generative models, particularly diffusion models, have shown promise in synthesizing realistic medical images, existing approaches face specific limitations:

Unconditional models often fail to capture rare pathological patterns.
Inpainting methods primarily focus on reconstructing healthy regions or performing super-resolution, rather than expanding images.
Latent Diffusion Models (LDMs), while standard for synthesis, struggle to recover high-frequency details, often producing blurry outputs.
Outpainting (expanding an image beyond original boundaries while preserving a specific region) remains underexplored in medical imaging, despite its potential to retain real pathological regions while generating diverse, surrounding healthy tissue.

Methodology

The authors propose POWDR (Pathology-preserving Outpainting with Wavelet Diffusion), a framework designed to generate full 3D MRI volumes that preserve real pathological regions while synthesizing anatomically plausible surrounding tissue.

1. Network Architecture and Wavelet Conditioning

POWDR builds upon a conditional wavelet diffusion framework to address the high-frequency detail loss common in LDMs.

Wavelet Decomposition: During the forward diffusion process, the original 3D MRI volume ( $y_0$ ) is decomposed using a digital wavelet transform (DWT) with a Haar kernel. This yields eight subbands ( $x_0$ ) representing low- and high-frequency components across different orientations (LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH).
Noise Addition: Gaussian noise is progressively added to the wavelet coefficients according to a linear $\beta$ -schedule.
Conditional Denoising: A 3D ResUNet architecture is employed for the denoising process. The conditioning image (masked pathology or random region) is also transformed into wavelet subbands and concatenated with the noisy target subbands, resulting in 16 input channels.
Loss Function: The network estimates the clean wavelet coefficients ( $\hat{x}_0$ ) from the noisy input, optimizing via Mean Squared Error (MSE) in the wavelet domain.
Sampling: During inference, the model uses 1,000 denoising steps (DDPM strategy). The generated latent sample is reconstructed into the image domain using the Inverse Discrete Wavelet Transform (IDWT).

2. Training Strategies

To prevent conditioning-induced collapse (where the model generates identical outputs for the same condition) and improve diversity:

Masked Pathology Conditioning: The primary strategy involves multiplying the pathology mask (e.g., tumor) with the intensity volume to serve as the condition.
Random Connected Mask Training: A secondary strategy introduces random connected masks (6-connected) sampled based on tumor volume distribution. These are multiplied by the image volume to serve as training conditions. This strategy is crucial for enhancing diversity in the non-pathological regions while maintaining the integrity of the conditioned area.

3. Datasets

The framework was evaluated on two distinct anatomical domains:

Brain MRI: Utilized the BraTS 2023 dataset (1,251 cases) for training and BraTS 2024 for evaluation. Images were normalized and cropped to $224 \times 224 \times 155$ voxels.
Knee MRI: Validated on a dataset comprising 768 volumes from public (DESS sequence) and internal (CUBE sequence) sources, including healthy and pathological cases (meniscus tears). Volumes were resampled to $288 \times 288 \times 160$ voxels.

Key Contributions

Novel Outpainting Framework: POWDR is presented as a method to condition on pathological regions to synthesize surrounding anatomy, enabling the generation of full MRI volumes with real, preserved pathology.
Comprehensive Evaluation: The work provides a rigorous assessment using quantitative metrics (FID, SSIM, LPIPS), diversity analysis via repeated sampling, and clinically relevant downstream tasks (tumor segmentation and tissue volume analysis).
Tissue-Agnostic Applicability: The method demonstrates extensibility beyond the brain, successfully applied to knee MRI, suggesting broad applicability across different anatomical regions.

Results

Quantitative Image Quality and Diversity

Realism: The generated images show high similarity to real data, with a Fréchet Inception Distance (FID) of 0.0042 (compared to 0.0005 for real-vs-real).
Pathology Preservation: High Multi-Scale Structural Similarity (MS-SSIM) of 0.9998 and low Learned Perceptual Image Patch Similarity (LPIPS) of 0.0030 were observed inside the tumor region, confirming the preservation of pathological features.
Anatomical Diversity: Outside the tumor, the model generated diverse structures (MS-SSIM 0.8195, LPIPS 0.1484).
Impact of Random Masks: The introduction of random connected mask training significantly improved diversity. Cosine similarity between pairwise samples dropped from 0.9947 to 0.9580, and KL divergence increased from 0.00026 to 0.01494, indicating a broader distribution of generated samples without compromising the conditioned pathology.

Clinical Relevance

Segmentation Performance: Using nnU-Net for tumor segmentation, adding 50 synthetic cases generated by POWDR improved the Dice score from a baseline of 0.6992 to 0.7137. Performance gains plateaued after 50 cases.
Anatomical Plausibility: Tissue volume analysis (CSF and Gray Matter) using FSL FAST showed no significant differences between synthetic and real images, validating the anatomical correctness of the generated volumes.
Knee MRI: The model successfully preserved meniscus tears while generating realistic surrounding anatomy in knee scans, despite variations in scanner vendors and sequences.

Significance and Claims

The paper posits that POWDR offers a practical solution to the data scarcity and class imbalance problems in medical imaging. By retaining real pathological regions and generating diverse, plausible surrounding tissue, the method creates a controllable framework for producing pathology-preserving synthetic data.

The authors emphasize that this approach:

Avoids the fabrication of lesions, ensuring that the pathological ground truth remains intact.
Mitigates the blurring issues associated with standard Latent Diffusion Models through wavelet-domain conditioning.
Is extensible to multiple anatomies (demonstrated on brain and knee) and supports robust model development for segmentation, classification, and vision-language tasks.

The work concludes that POWDR is a viable strategy for augmenting medical datasets, particularly for low-prevalence conditions, without incurring additional acquisition costs.