In critical emergencies, such as disasters, accidents or conflict zones, unidentified patients frequently present in hospital settings. The inability to determine a person’s identity promptly can hinder medical decision-making and delay communication with relatives. While computed tomography (CT) is primarily used for diagnostic purposes, its routine availability presents an opportunity for a secondary application: personal identification. A novel method has been developed that uses computer vision (CV) to match thoracic CT scans, specifically maximum intensity projection (MIP) images, to a reference database. This approach leverages unique anatomical structures within the chest to reliably identify individuals, even without access to traditional identifiers.
Thoracic CT as a Distinctive Identifier
Traditionally, attempts at personal identification using CT have focused on axial slices of the head. These methods, however, are limited by their susceptibility to changes in head orientation and artefacts caused by metallic objects. The thoracic region offers a more stable alternative. The sternum, spine and thoracic skeleton as a whole present highly individualised anatomical features that remain relatively constant over time. The study in question created 2D MIP images by selecting the brightest pixel from each slice in a 3D CT volume. These images, although lacking in contrast initially, were enhanced using contrast-limited adaptive histogram equalisation (CLAHE), a method that improves image visibility by adjusting contrast locally within small tiles.
Must read: Enhancing Emergency Radiology with Retrieval-Augmented AI
Once optimised, the images were analysed using the AKAZE algorithm, which identifies keypoints—distinct areas such as corners or edges—and encodes their local features into descriptors. These descriptors are invariant to changes in scale and rotation, making them highly suitable for reliable matching across different scans. Matching is carried out by comparing descriptors from a query image to those in a database using a Euclidean distance metric. The closest and second-closest matches are then filtered through Lowe’s ratio test to retain only the most reliable matches. The random sample consensus (RANSAC) algorithm is used to eliminate outliers, ensuring only consistent, high-quality matches are retained. A final score is generated, reflecting the proportion of matching features between the images.
Robust Identification Across Populations
The retrospective study analysed 12,465 thoracic CT examinations from 8177 patients. A total of 300 identification trials were conducted using MIP images from this cohort. The results were remarkable: 296 of the 300 identifications were accurate at rank 1 (the highest match), with only four cases requiring further ranks for correct identification. By rank 10, identification accuracy had reached 99.67%. The median similarity score for images of the same person was substantially higher than for different individuals, affirming the method’s precision. For example, the average score for correct matches was 7.43%, compared to just 0.16% for incorrect matches.
Notably, the method proved effective across all age groups, with minor score reductions observed in older individuals. These were mainly due to factors such as reduced image contrast or the presence of medical equipment, which can obscure key anatomical features. Nonetheless, even when CT scans were taken several years apart, the method consistently identified the correct individual, underscoring its robustness. The approach performed well without needing to compare identical scans or limit itself to short time intervals between acquisitions.
An essential component of this success was parameter optimisation. Systematic variation of key settings—such as CLAHE intensity (clipLimit), tile size, AKAZE feature extraction parameters and matching thresholds—revealed that the best results were achieved when contrast enhancement was applied and the matching process was run in both directions (query to reference and vice versa). This ensured consistent outcomes, even when query images differed in orientation or detail.
Privacy, Practicality and Future Directions
While this CV-based identification method is highly accurate, it is not intended to serve as a legally binding means of identification. Rather, its primary value lies in narrowing down the pool of potential identities, enabling medical professionals to make quicker, more informed decisions. Once a probable match is found, a formal forensic confirmation can follow using established legal protocols.
Importantly, the system stores only CV features—abstract numerical representations of local image areas—without retaining the original CT scans. These features are decoupled from the image data and contain no personally identifiable information. Patient metadata such as ID, age and sex is pseudonymised and encrypted, preserving confidentiality. This approach could be extended by maintaining centralised databases of CV features that include only the origin of the scan. When a high match is detected, the original institution can be queried for identity confirmation, reducing the risk of privacy violations.
Nevertheless, challenges remain. One is the optimal selection of CT slices for MIP image generation. Including too many slices may introduce irrelevant data or reduce contrast. Future studies could examine whether selecting only the most informative slices enhances accuracy. Additionally, applying demographic filters—such as age, sex or body size—could further refine results. Research into how trauma, ethnicity or postmortem changes affect performance is also warranted. Moreover, comparisons between thoracic MIP images and conventional X-rays could broaden the method’s clinical utility, particularly in resource-limited settings.
There is also potential to use postmortem CT images in forensic contexts, matching them to antemortem records for the identification of deceased individuals. This application could be particularly useful in mass casualty events or among displaced populations.
The use of CV-based analysis of MIP images from thoracic CT scans represents a significant advancement in the field of personal identification, especially in emergency medicine. The method has demonstrated exceptional accuracy, even across long time intervals and varied patient demographics. By harnessing the unique anatomical features of the thoracic skeleton and combining them with sophisticated image processing and feature-matching algorithms, it offers a highly practical tool for use in high-pressure scenarios. While not a substitute for formal identification procedures, it provides a crucial first step in narrowing down possible identities, ultimately enhancing patient care and operational efficiency. As further refinements are developed and ethical safeguards implemented, this technique may become a cornerstone of emergency identification protocols in hospitals around the world.
Source: European Radiology
Image Credit: iStock
