E5152. Validation of a Deep Learning Algorithm in Detecting Acute Cervical Spine Fractures on CT
Authors
Nicholas Manasewitsch;
University of Washington
Kevin Chorath;
University of Washington
Nitin Venugopal;
University of Washington
Aman Buttar;
University of Washington
Akinpelu Babatunde;
University of Washington
Mahmud Mossa-Basha;
University of Washington
Objective:
Cervical spine fractures are devastating injuries that may result in permanent disability, and prompt diagnosis is essential to ensuring the best possible patient outcomes. Deep learning algorithms are critical tools that can assist in triaging patients with specific acute findings to optimize worklist management and enhance patient care. This study was designed to analyze the diagnostic accuracy and radiologist concordance of a deep learning algorithm, specifically Aidoc’s cervical spine tool, in detecting acute cervical spine fractures on CT.
Materials and Methods:
A retrospective study was performed with 2663 noncontrast and contrast-enhanced CT studies performed from November 2020 to March 2021 that included the cervical spine. CT studies of the head, neck, and spine were included. The patient population included trauma and nontrauma adult patients in the emergency room setting at a level 1 trauma center. Aidoc’s interpretation of the presence or absence of acute cervical spine fractures was compared to the final attending report, and discrepant cases were independently analyzed by attending neuroradiologists.
Results:
The prevalence of acute cervical spine fracture in our sample was 4.3%. The concordance between the attending report and Aidoc’s interpretation was 98.3%. Aidoc correctly identified 90 out of 108 acute cervical spine fractures with 28 false-positive cases. In our patient population, the sensitivity was 83.3%, and specificity was 98.9%. Examples of reasons for false-positive cases identified by Aidoc included incorrectly identifying atherosclerotic calcifications, chronic spine fractures, beam artifact, and sites of prior hardware removal as acute cervical spine fractures.
Conclusion:
Our study demonstrated moderate sensitivity and high specificity in the deep learning algorithm’s ability to identify acute cervical spine fractures. Though further analysis and external validation is required, deep learning algorithms, such as Aidoc, are emerging tools that could assist in triaging cases and worklist-management by prioritizing flagged cases.
