Fingerprint: An AI-Based Method For Detection of Mislabeled CT Studies In Clinical Trials
Author(s): Lars Edenbrandt, Olof Enqvist, Tony Gillberg.
The manual step of replacing the identification number of each subject in clinical trials with the correct trial identification number is critical. Mislabeling studies may lead to excluding subjects or visits or to false study results. This article presents an automated AI-based method to detect if a pair of de-identified CT studies has been obtained from the same subject. Approach: An automated segmentation of bones in CT is performed using Organ Finder (SliceVault AB, Malmö, Sweden). Corresponding bones from the two images are aligned using Iterative closest point. The percentage of vertices in the smaller mesh with residuals below 1.5 mm is used as a measure of similarity, the anatomical match. If the anatomical match is less than 80% for the left and right hip bone and the left and right scapula, the two CT studies are classified as being obtained from different subjects. The Fingerprint method was tested on a group of 58 patients, who each had two CT studies obtained at different occasions. From the 116 CT studies 6,612 unique pairs of CT from different patients and 58 pairs from the same patient could be selected. Results: The Fingerprint method classified all 6,612 pairs of CT studies from different patients correctly (sensitivity 100%; 95% confidence interval 99-100%) and all 58 pairs of CT studies from the same patients correctly (specificity 100%; 95% confidence interval 94- 100%). Conclusion: This study shows how an AI-based method can be used to accurately detect if a pair of de-identified CT studies has been obtained from two different subjects.