Anonymizing Radiographs Using an Object Detection Deep Learning Algorithm

• DOI: 10.1148/ryai.230085
• Publication Date: 2023-09-13T13:51:28Z
• Authors: Bardia Khosravi; John P. Mickley; Pouria Rouzrokh; Michael J. Taunton; A. Noelle Larson; Bradley J. Erickson; Cody C. Wyles

Radiographic markers contain protected health information that must be removed before public release. This work presents a deep learning algorithm localizes radiographic and selectively removes them to enable de-identified data sharing. The authors annotated 2000 hip pelvic radiographs train an object detection computer vision model. Data were split into training, validation, test sets at the patient level. Extracted then characterized using image processing algorithm, potentially useful (eg, "L" "R") without identifying retained. model achieved area under precision-recall curve of 0.96 on internal set. de-identification accuracy was 100% (400 400), with false-positive rate 1% (eight 632) retention 93% (359 386) for laterality markers. further validated external dataset chest radiographs, achieving 96% (221 231). After fine-tuning 20 images from investigate potential improvement, 99.6% (230 231, P = .04) decreased 5% (26 512) achieved. These results demonstrate effectiveness two-pass approach in de-identification. Keywords: Conventional Radiography, Skeletal-Axial, Thorax, Experimental Investigations, Supervised Learning, Transfer Convolutional Neural Network (CNN) Supplemental material is available this article. © RSNA, 2023 See also commentary by Chang Li issue.