Braineditor: Structure-Disentangled Brain MRI Synthesis via Natural Language Prompted Diffusion Image Editing

Deep learning has advanced medical image synthesis models to tackle data shortages and assist in disease prediction and analysis. In this work, we introduce BrainEditor, a novel approach to structure-disentangled brain MRI synthesis via text-prompted diffusion image editing. Unlike most existing meth-ods that often alter multiple anatomical structures in a linked manner, our approach allows precise control over individual structures. This enables the simulation of potentially diverse and complex disease patterns in clinical practice. By integrating textual prompts into the image synthesis process, we en-hance BrainEditor's flexibility for user interaction. To achieve the disentangled synthesis, we generate artificially structure-disentangled training image pairs through Poisson image editing’ and link text descriptions to image progressions by quan-tizing volumetric change rates of brain structures into limited intervals. Experiments on the public OASIS-2 dataset show that BrainEditor accurately synthesizes high-quality images following textual instructions and improves the performance of downstream tasks through data augmentation. The code is available at https://github.com/LIKPO/BrainEditor.