Chest X-rays: a diffusion editor modifies a specific area without affecting the rest of the image in 93.5 per cent of cases

Tested on 2,400 requests drawn from the MIMIC dataset, the system combines a learned gate and a mask to apply the requested modification whilst preserving the regions not affected.

Generating controlled variants of an X-ray (adding or removing an anomaly, for example) is useful for testing diagnostic AIs or enriching datasets, as long as only what is requested is modified. Kamran Ullah Afaq and Basit Raza present on arXiv a latent diffusion editor that evolves two trajectories in parallel: one reconstructs the source image independently of the instruction, the other applies the modification. A learned door mixes the two at each step, then a mask in the image space glues the proposal back to the original.

Gains measured component by component

Out of 2,400 requests built from the MIMIC database, 2,244 results satisfied the target, preservation, quality and coverage, i.e. 93.5%, and the requested modification was made in 97.4% of cases. The system is 1.7 points higher than a version limited to regluing, with the learned door and the mask each providing a measurable gain. The error in the protected areas fell from 0.0190 to 0.0075 after regluing, and a blinded evaluation of 240 requests confirmed the discrepancy (93.3% versus 91.7%).