https://atlas.rsna.org/schemas/2025-11/model.json

Anonymizing Radiographs Using an Object Detection Deep Learning Algorithm

https://doi.org/10.1148/ryai.230085

1.0

Cody C. Wyles (email: ude.oyam@ydoC.selyW)

Supported by the Mayo Foundation Presidential Fund.

Institutional review board approved with waiver of informed consent.

Two-pass pipeline: (1) YOLOv5-x (Ultralytics) object detection model to localize radiographic markers; (2) postprocessing to retain desired markers (e.g., laterality) using Otsu thresholding, connected component analysis, orientation correction, and OCR (Tesseract with LSTM) to detect isolated R/L; non-retained components are replaced by black pixels.

Open-source code: https://github.com/OrthopedicSurgeryAILab/RadiographMarkerRemover

Enables de-identification of radiographs by removing burnt-in markers containing PHI while selectively retaining useful non-identifying markers (e.g., laterality), facilitating data sharing and robust AI model development.

Data de-identification and preprocessing for research data sharing and AI development.

Fully automated detection and removal of radiographic markers with optional automated retention of specified markers (e.g., R/L).

Automated anonymization of radiographs by localizing and removing radiographic markers containing protected health information; intended for radiographic images (e.g., pelvis, chest) in research data pipelines and multi-institutional data sharing environments.

Radiographic images (e.g., anteroposterior pelvis and chest radiographs).

Run the YOLOv5-x localizer to detect all marker regions; apply the postprocessing/OCR module to retain desired markers (e.g., isolated R/L) and redact others; optionally fine-tune the localizer on a small sample (e.g., 20 images) from a target domain to improve generalization; for training downstream AI models, disable marker retention to avoid shortcut learning from markers.

External public images lacked PHI, so external evaluation focused on study- and staff-related markers; algorithm may require fine-tuning to cover other marker types and other body parts/modalities; false positives occurred with non-marker items (e.g., ECG leads) not seen in training; OCR-based retention can be sensitive to image quality and marker variability.

For deployment on new domains, fine-tune with a small set of local images; retain only strictly non-identifying markers as needed; disable retention during training of diagnostic models to prevent shortcut learning.

Annotations created with LabelMe v5.0.1; model trained in PyTorch v1.11.0; OCR with Tesseract v5.0.0; code and pipeline available on GitHub with details in Appendix S1.

Average processing time per image: 1.43 s on CPU (Intel i7 2.5 GHz) and 0.51 s on GPU (NVIDIA V100).