Automated Medical Image De-Identification via U-Net++ Segmentation and Conditional GAN Inpainting

The acceleration of multi-centric medical AI studies hinges on the ability to share imaging data without exposing burnt-in Protected Health Information (PHI). Manual redaction remains the dominant practice, but it erases diagnostically relevant context, violates harmonization guidelines issued by large consortia, and cannot keep up with the petabyte-scale repositories envisioned by regulatory agencies. This study delivers a comprehensive treatment of a fully automated Detect-and-Restore pipeline that fuses fine-grained U-Net++ segmentation with a context-aware conditional GAN (cGAN) inpainter. Building on two engineering notebooks (U-Net++ training and GAN generator orchestration), we develop a synthetic PHI rendering engine, a dynamic oracle that freezes the detector during adversarial optimization, and a hybrid loss that couples adversarial, pixelwise, and perceptual cues. Extensive experiments on 48,000 synthetically annotated radiographs demonstrate a Dice score of 0.8147 for PHI localization and a PSNR/SSIM/LPIPS triplet of 41.87 dB/0.985/0.027 for restoration while keeping inference below 92 ms per image on a single RTX 4090. Beyond reporting raw metrics, we dissect error modes, quantify the effect of imperfect masks on the inpainter, and position the proposal relative to recent international initiatives on medical image de-identification. Testing on an external clinical cohort of 200 real-world DICOM radiographs confirms generalizability, maintaining a PSNR of 40.12 dB and demonstrating robust blending at masking boundaries without compromising downstream diagnostic utility across heterogeneous hospital data.