Anonymizing Radiographs and Making Them “AI Ready” Using a Deep Learning Algorithm

Purpose: Radiographs are the most frequently performed radiologic study given their low cost, minimal radiation dose, and diagnostic richness. Most radiographs have radio-opaque markers to denote the side, imaging unit, and patient information (such as medical record number), often set in place by technicians, causing great variation between studies. In this work, we developed a Deep Learning algorithm to detect and remove radiologic markers to enable de-identified data sharing, and also make the images more suitable for a robust DL training pipeline.

Methods/Materials: We annotated 2000 hip/pelvis radiographs to create our ground-truth dataset. During annotation, each marker area was denoted by a bounding box. Then, the data was split at the patient level into 60%/20%/20% for train/validation/test sets. A YOLOv5-x model was trained on 512x512 pixel images to detect the marker location. Mean average precision at a 50% threshold (mAP-50) was reported as a measure of model performance on the test set. To externally validate the model, we used the validation set of the CheXpert dataset and applied the model on chest radiographs to evaluate performance on data from alternate anatomical regions and different organizations. Through backbone freezing, the model was further fine-tuned on 20 images from the CheXpert training set to investigate the potential for improvement in target task performance. Annotations were executed using the LabelMe software (v5.0.1), and models were trained using PyTorch (v1.11.0).

Results: The model reached an mAP-50 of 0.99 on the internal test set of hip and pelvis radiographs (Figure 1). On the external test set of chest radiographs, the model completely removed markers from 96% of radiographs (222/231). After fine-tuning on 20 CheXpert radiographs, the model was able to remove markers from 100% (231/231) of the chest radiographs.

Conclusions: In this work, we have developed a model that can accurately remove radiographic markers from various radiographs. With minimal effort, this model can be finetuned and be used to effectively detect radiographic markers out of the training set distribution, highlighting its generalizability.