A new diffusion model can erase a tumor from a brain MRI scan and paint in what the tissue probably looked like before it grew there.
The tool, called CATCH, works in the Haar-wavelet domain rather than directly on pixels, which its creators say helps it reconstruct texture while a hard-compositing step leaves every real voxel outside the tumor mask untouched. The team tried three ways of shaping the training masks: fixed masks, tumor-shaped augmentation, and a weighted mix of tumor-derived, blob-shaped, and ellipsoidal masks. After comparing frozen versions of all three on a 75-case internal set, they picked the weighted mixture and submitted only that version to the BraTS 2026 brain-tumor challenge for official grading. On the challenge's 219-case validation set, that model scored an SSIM of 0.772, a PSNR of 20.89dB, and an MSE of 0.0098 - roughly in line with the 0.80 SSIM and 19.18dB PSNR it hit on the researchers' own internal test averaging.
That's a solid result for a tricky task: convincingly hallucinating healthy anatomy where a tumor used to be, without smearing the surrounding brain. The catch is what the challenge didn't test. The researchers also ran an internal comparison against a simpler, compute-matched random-augmentation baseline and found real, statistically significant gains in SSIM, PSNR, and error after correcting for multiple comparisons - but that comparison never touched the official 219-case validation set, and neither did the fixed-mask or random-augmentation versions on their own.
Useful research, honestly reported limits - the paper says as much - but until fixed and random baselines get their own official scores, or someone runs an external comparison, saying CATCH beat the alternatives holds up only inside the authors' own test set.