⚡ electrical engineering

$\texttt{DisMorph}$ : learning to disentangle technical distortions from true biological change

The paper introduces $\texttt{DisMorph}$ , a synthetic-data-trained registration framework that disentangles technical MRI distortions from true biological brain changes to enable more accurate longitudinal morphometry in neurodegenerative disease studies.

Original authors: Jingru Fu, Kathleen E. Larson, Douglas N. Greve, Bruce Fischl, Malte Hoffmann

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jingru Fu, Kathleen E. Larson, Douglas N. Greve, Bruce Fischl, Malte Hoffmann

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 measure how much a tree has grown over a year. You take a photo today and another photo next year. If you could magically freeze the camera and the lighting, comparing the two photos would be easy; you'd just see the new branches. But what if, between the two photos, the camera lens warped slightly, making the tree look squished on the left in the second picture? Or what if the lighting changed, casting a shadow that made the trunk look thinner? You might mistakenly think the tree shrank, when in reality, it just looked different because of the camera.

This is the exact problem scientists face when studying the human brain using MRI scans. MRI machines are like incredibly powerful cameras that take 3D pictures of our brains. Doctors and researchers use these pictures to track how the brain changes as we age or as diseases like Alzheimer's take hold. They look for "atrophy," which is the natural shrinking of brain tissue, much like a tree losing leaves. However, the MRI machines themselves aren't perfect. They introduce "distortions"—tiny, invisible warps in the image caused by the machine's magnets and electronics. These warps can make a healthy brain look like it's shrinking, or hide the fact that a sick brain is shrinking faster than it should. For years, scientists have tried to fix these machine errors, but often the "fix" isn't perfect, leaving behind a messy mix of real brain changes and fake machine errors.

This is where a new tool called DisMorph comes in. Think of it as a super-smart detective that can look at two brain scans taken years apart and say, "Okay, this part of the change is because the brain actually got smaller, but this other part is just the camera lens warping the picture."

The researchers behind DisMorph, led by Jingru Fu and colleagues, realized that existing methods were trying to solve the whole puzzle with one big guess. They would calculate a single "deformation" map that tried to explain every difference between the two scans. The problem is, this single map mixes the real biological changes (the tree growing or shrinking) with the technical glitches (the camera warping). If you don't separate them, your measurements of brain health can be wrong.

To fix this, the team built a system that learns to untangle these two effects. They didn't just feed the computer real brain scans; instead, they taught it using a "synthetic" training ground. Imagine a video game where the computer generates millions of fake brain scans. In this game, the computer deliberately adds two things: a specific amount of "atrophy" (simulating a shrinking brain) and a specific amount of "distortion" (simulating the MRI machine's warping). Because the computer knows exactly how much of each it added, it can learn to spot the difference. It's like training a dog to distinguish between a real rabbit and a stuffed toy rabbit by showing it thousands of examples where it knows the answer.

Once the system, DisMorph, was trained on these fake scenarios, the researchers tested it on real data. First, they looked at scans where the only difference was the machine's distortion (no actual brain change). DisMorph successfully identified that almost all the change was just the machine warping the image, leaving the "brain change" part at zero. This proved it could tell the difference between a glitch and a real event.

Next, they tested it on real patients with Alzheimer's disease, looking at scans taken two years apart. In these cases, the brain should be shrinking. DisMorph successfully separated the two effects again. It identified the smooth, widespread warping caused by the scanner as "technical distortion" and the specific, localized shrinking of brain structures (like the hippocampus, which is crucial for memory) as "true biological change."

The results showed that DisMorph is better at finding the real brain changes than the old methods. In simulations where they knew the exact answer, DisMorph was more accurate. In real-world tests, it found that even after standard corrections, some machine distortion often remains, which could have been hiding or faking brain changes in previous studies.

The paper suggests that by using this new method, scientists can get a clearer, more honest picture of how the brain changes over time. It doesn't claim to have solved every problem in MRI imaging, but it offers a powerful new way to separate the signal (the real brain) from the noise (the machine's quirks). This is especially important for studies that look back at old data or compare scans from different hospitals, where the "camera lenses" might be slightly different. By learning to disentangle the technical mess from the biological truth, DisMorph paves the way for more accurate measurements of aging and disease, helping us understand the brain's journey a little better.

Technical Summary of DisMorph: Learning to Disentangle Technical Distortions from True Biological Change

Problem Statement
Longitudinal MRI is essential for quantifying structural brain changes associated with aging and neurodegenerative diseases. However, accurate measurement is compromised by geometric distortions introduced by MRI hardware, particularly gradient non-linearity (GNL). These technical distortions vary across scanners and acquisition protocols. Conventional deformable image registration methods estimate a single deformation field to align longitudinal scans, which conflates true biological changes (e.g., atrophy) with technical artifacts. This conflation can bias downstream morphometric measurements, reducing statistical power and obscuring disease progression, especially when residual distortions persist despite vendor corrections or when retrospective datasets lack consistent correction protocols.

Methodology
The authors propose DisMorph, a learning-based registration framework designed to explicitly decompose longitudinal deformation into two distinct components: a technical distortion field and an anatomical change field.

Synthetic Training Strategy: DisMorph is trained entirely on synthetic data generated via a novel two-stage generative process to avoid reliance on real-world ground truth labels:
1. Geometry Generator: Starting from anatomical label maps, this module applies a smooth random elastic transform to simulate GNL distortion. It then introduces anatomical changes (atrophy or growth) by iteratively reassigning boundary voxels between target structures and neighbors based on sampled volumetric change ratios.
2. Intensity Generator: This module synthesizes realistic grayscale images from the distorted and atrophied label maps, applying random intensity means, variances, and corruptions (bias fields, gamma transforms, blurring, noise) to ensure robustness to real-world acquisition variability.
Network Architecture: The model employs a dual-stream encoder with shared weights to process fixed and moving images. It utilizes a pyramid-warping encoder to predict two separate stationary velocity fields (SVFs): one for distortion ( $v_d$ ) and one for anatomical change ( $v_a$ ). These are integrated via scaling and squaring to produce diffeomorphic deformation fields ( $\phi_d$ and $\phi_a$ ).
Training Objective: The model is supervised using the known label maps from the synthesis process. The loss function minimizes the mean squared error between the warped label maps and the ground truth labels for both the distortion and anatomical components, combined with a smoothness regularization term.

Key Contributions

Explicit Disentanglement: DisMorph is the first registration framework to explicitly separate longitudinal deformation into technical (distortion) and anatomical (atrophy/growth) transforms, rather than estimating a single conflated field.
Synthetic-Only Training: The framework utilizes a generative model that synthesizes both geometric effects and intensity variations independently, providing "disentanglement supervision" without requiring real-world paired data with known ground truth.
Domain Randomization: The training pipeline incorporates domain randomization across imaging protocols and contrasts to promote generalization across different scanners and acquisition systems.

Experimental Results
The authors evaluated DisMorph in three settings:

Simulated Data (Known Ground Truth): Using biomechanical atrophy models with prescribed volume changes, DisMorph demonstrated significantly lower Atrophy Estimation Error (AAE) compared to conventional "standard" registration (which uses a single composite field). The improvement was most pronounced in low-atrophy scenarios, corresponding to early disease stages.
Distortion-Only Real Data (fBIRN): On image pairs differing only by GNL distortion (original vs. vendor-corrected), DisMorph correctly attributed the vast majority of geometric change to the distortion field ( $\phi_d$ ), leaving the anatomical field ( $\phi_a$ ) with minimal signal. This confirmed the model's specificity in the absence of biological change.
Longitudinal Alzheimer's Disease Data (ADNI): On two-year longitudinal pairs from AD patients, DisMorph successfully isolated the expected subcortical atrophy in the anatomical field while identifying residual distortion in the distortion field. Visual inspection confirmed that the model preserved localized tissue loss (e.g., hippocampal atrophy) while correcting for smooth, scanner-induced misalignments.

Significance and Claims
The paper claims that by disentangling MRI-induced distortion from biological change, DisMorph enables more accurate longitudinal morphometry, particularly in clinical settings where maintaining acquisition consistency is challenging or where retrospective data may have uncorrected or inconsistently corrected distortions. The authors note that their analysis of ADNI data suggests residual scanner-induced distortion persists even after standard vendor corrections. They position DisMorph as a complementary tool to vendor corrections, capable of improving the sensitivity of disease progression studies by reducing distortion-related bias. Future work is proposed to extend this decomposition to other distortion sources (e.g., B0 inhomogeneity) and to investigate the impact of this decomposition on downstream statistical models of disease.

More like this