End-to-end deep learning pipeline redacts protected health information from medical images and inpaints anatomy to preserve downstream analysis

A deep learning pipeline detected burned-in text in medical images with F1=0.891 (recall 0.912, precision 0.875) and used generative inpainting to restore redacted regions. Downstream segmentation Dice scores remained comparable to originals, ranging from 0.936 to 0.959 across i…

• The framework combines a CRNN-based redaction module with a latent-diffusion inpainting module (Stable Diffusion 2).
• De-identified images maintained high fidelity for anatomy segmentation, with Dice ranging 0.936–0.948 on one dataset and 0.955–0.959 on another.

Automated summary

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