# Multicenter, Multinational, and Multivendor Validation of an Artificial Intelligence Application for Acute Cervical Spine Fracture Detection on CT

## TL;DR

This study shows that an AI tool can accurately detect acute cervical spine fractures using CT scans from diverse global sources.

## Contribution

The study validates AI performance for cervical spine fracture detection across multiple centers, countries, and CT scanner vendors.

## Key findings

- The AI achieved 90.3% sensitivity and 91.9% specificity in detecting acute cervical spine fractures.
- The AI correctly localized fractures in 84.4% of bounding boxes and labeled spinal levels with 97.3% accuracy.

## Abstract

Background/Objectives: While previous studies have evaluated AI algorithms for cervical spine fracture (CSFx) detection on CT, many have lacked validation on diverse, multinational datasets or have focused primarily on overall case-level classification This study aimed to evaluate the performance of an AI application for acute CSFx detection in case-level classification, fracture localization, and spinal level labeling on multicenter, multinational, and multivendor CT data. Methods: Non-enhanced CTs were retrospectively collected from a U.S. teleradiology company, a French teleradiology company, and a U.S. university hospital. Four radiologists independently labeled the presence and location (including the spinal level) of acute CSFx to establish the reference standard. Per-case diagnostic performance, per-bounding box positive predictive value (PPV) for localization, and overall agreement of cervical vertebral level labeling of the AI were assessed. Results: A total of 155 patients (60.6 years ± 21.2 years, 104 men) with acute CSFx and 173 patients (51.9 years ± 22.7 years, 91 men) without acute CSFx were evaluated. Data were acquired using scanners from five manufacturers. For acute CSFx diagnosis, the AI achieved a per-case sensitivity of 90.3%, a specificity of 91.9%, an accuracy of 91.2%, an area under the receiver operating characteristic curve (AUC) of 0.91, and Matthews correlation coefficient of 0.82. Among 192 bounding boxes representing acute CSFx generated for 154 positive cases by the AI, 162 were true positives (per-bounding box PPV, 84.4%). Of the 186 bounding boxes for which the AI displayed cervical spinal level, 181 were labeled correctly (overall agreement, 97.3%). Conclusions: The AI application for detecting acute CSFx demonstrated high diagnostic performance on multicenter, multinational, and multivendor data, with high performance in fracture localization and spinal level labeling.