It fills in diseased brain scan regions with healthy tissue without changing the rest.

Diseased brain tissue can disrupt automated medical scans. Researchers want to fill these damaged areas with realistic healthy tissue, a process called inpaintinginpaintingFilling in missing or corrupted parts of an image.. To solve this, the authors created RARF. Instead of changing the entire three-dimensional scan, RARF leaves the healthy areas untouched to preserve the patient's unique anatomy. It only works inside the missing zone, which it initially fills with random Gaussian noise. A neural network learns to gradually turn that noise into realistic brain tissue over time using learned trajectories called rectified flows. Finally, the system merges the newly created tissue with the original, unchanged brain. Tests following the BraTS protocol showed that RARF generates competitive reconstructions while keeping the anatomy consistent with the rest of the brain.

Why it matters

For doctors and AI systems analyzing brain scans, this could make it easier to study healthy tissue patterns in patients with brain tumors. It provides complete images even when parts are damaged, potentially improving diagnosis and treatment planning.

Questions answered