strong-followup
MedEdit: Counterfactual Diffusion-based Image Editing on Brain MRI
Malek Ben Alaya, Daniel M. Lang, Benedikt Wiestler · Technical University of Munich · 2024-07 · arXiv:2407.15270
Why it matters
Induces pathology into healthy brain MRI while preserving the surrounding scan. Notable for its evaluation: a board-certified neuroradiologist judged the generated strokes indistinguishable from real ones, which is a stronger bar than any distribution metric.
What this paper does
read: full textBefore this
Counterfactual biomedical editing had mostly targeted chest X-ray disease progression using paired longitudinal data, and separate brain-imaging work generated healthy scans from pathological ones for lesion localization. Diffusion editing tools like SDEdit and Palette existed for natural images, but the paper states existing methods "either inadequately model indirect pathological effects like brain atrophy or by excessively altering the scan, which disrupts correspondence to the original images."
The problem
Induce a stroke lesion into a specific region of a brain MRI, allowing plausible indirect effects such as atrophy nearby, without disturbing the rest of the scan's correspondence to the original.
The idea
Extend RePaint-style diffusion inpainting from class-conditional to mask-conditional generation, and inpaint a deliberately dilated region around the target pathology rather than only the pathology mask itself, so the model has room during denoising to generate secondary tissue changes the true lesion would cause.
How it works
MedEdit operates in pixel space. A standard DDPM U-Net denoises 128x128 T1-weighted brain slices directly, with no VAE, VQGAN, or other latent compression anywhere in the pipeline. The paper gives no explicit justification for pixel space; it is presented as the standard DDPM setup, T equals 1000 steps with a linear beta schedule from 1e-4 to 0.02, without discussing the representation choice. This means MedEdit does not use a borrowed Stable Diffusion latent the way its line categorization in this atlas otherwise implies; the mechanism here is pixel-space RePaint-style inpainting with a dilated conditioning mask, not fine-tuned Stable Diffusion with a frozen natural-image VAE. The core technical contribution is the mask dilation, kernel size 25, applied with four resampling iterations per timestep to harmonize the inpainted and known regions.
Evidence
The paper offers the strongest evidence tier in this batch, a downstream segmentation task and a clinician reader study, alongside distribution metrics. On the Atlas v2.0 stroke dataset, 655 T1 scans with 54 held out for test, MedEdit reaches FID 3.07 against Palette's 5.63 and SDEdit's 7.95, a 45% and 61% improvement respectively, and a Dice score of 0.63 from an nnUNet lesion segmenter run on the generated counterfactuals, against Palette's 0.38 and naive RePaint's 0.50. Beyond distribution metrics, a board-certified neuroradiologist rated 80 generated and 20 real scans, blinded, on a 1 to 5 scale. MedEdit's realism score of 3.20 matched the real scans' own 3.20 and beat SDEdit's 2.80, and MedEdit scored 3.15 on indirect pathological realism against SDEdit's 3.00, Palette's 2.00, and naive RePaint's 1.85. This is a genuine reader-study result, not just a metric, though it rests on a single rater with no inter-rater reliability reported.
Limitations
The paper concedes its own computational metrics "fall short in capturing nuanced clinical realities" and that FID specifically "overlooks critical subtleties like the indirect pathological effects accompanying stroke lesions," which is why it added the reader study. It concedes the method was evaluated only in 2D at 128x128 and proposes 3D extension as future work. It concedes the neuroradiologist evaluation used a single rater on a subjective scale without inter-rater reliability. It concedes the method "was not designed specifically for stroke synthesis" and claims generalization to other pathologies without evidence for it.
Why it matters
It shows that a reader study can move a medical generation paper's central claim, indistinguishability from real, onto ground a distribution metric cannot reach, and it shows this is achievable with the plainest possible generation space, plain pixel-space DDPM, no latent compression at all.
Abstract, in the authors' own words
Denoising diffusion probabilistic models enable high-fidelity image synthesis and editing. In biomedicine, these models facilitate counterfactual image editing, producing pairs of images where one is edited to simulate hypothetical conditions. For example, they can model the progression of specific diseases, such as stroke lesions. However, current image editing techniques often fail to generate realistic biomedical counterfactuals, either by inadequately modeling indirect pathological effects like brain atrophy or by excessively altering the scan, which disrupts correspondence to the original images. Here, we propose MedEdit, a conditional diffusion model for medical image editing. MedEdit induces pathology in specific areas while balancing the modeling of disease effects and preserving the integrity of the original scan. We evaluated MedEdit on the Atlas v2.0 stroke dataset using Frechet Inception Distance and Dice scores, outperforming state-of-the-art diffusion-based methods such as Palette (by 45%) and SDEdit (by 61%). Additionally, clinical evaluations by a board-certified neuroradiologist confirmed that MedEdit generated realistic stroke scans indistinguishable from real ones. We believe this work will enable counterfactual image editing research to further advance the development of realistic and clinically useful imaging tools.
Research line
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.