Controllable panoramic radiograph generation with PRGen

PRGen, a disease-guided generative model trained on 50,127 text-image pairs, synthesizes controllable panoramic radiographs. Adding generated images to training boosted segmentation Dice 47.59% and AUC 11.53%; >82% of images deemed clinically realistic. External validation showe…

• Trained on 50,127 paired text-image samples for controllable generation.
• Enables control via textual disease descriptions or simple/detailed sketches.
• External multicenter validation confirmed robust generalizability with 25.58% mean Dice improvement for segmentation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.