E2651. Using Deep Learning to Detect Atlantoaxial Instability
  1. David Lee; David Geffen School of Medicine at UCLA
  2. Ming-Yeah Hu; David Geffen School of Medicine at UCLA
  3. Raffi Salibian; Olive View-UCLA Medical Center; David Geffen School of Medicine at UCLA
Anteroposterior atlantoaxial instability (AAI) is characterized by increased mobility between the atlas (C1) and the axis (C2) in the anteroposterior plane. AAI may result from both osseous and ligamentous abnormalities at C1-C2 and can be seen with congenital conditions like Down syndrome or may be acquired after trauma or an inflammatory arthritis, like rheumatoid arthritis. A plain radiograph of the cervical spine in lateral projection is the most commonly used imaging modality to diagnose AAI. Flexion-extension views of the cervical spine in lateral projection can provide better diagnostic accuracy in AAI. The atlantodental interval (ADI) is measured between the anterior border of the dens and the posterior border of the anterior arch of C1. An ADI > 3 mm in adults is suggestive of AAI. We aim to develop a deep learning platform to automatically flag radiographs concerning for AAI. Automated detection of AAI can assist Radiologists with workflow optimization to ensure cases of AAI are interpreted in a timely manner. Physicians in clinical specialties, such as emergency medicine and rheumatology, may also benefit from this model when radiologist expertise is scarce or not immediately available.

Materials and Methods:
After obtaining Institutional Review Board (IRB) approval, we surveyed the available adult cervical spine radiographs from 2007 to 2022 at our institution. We measured the ADI between the posterior border of the atlas and the anterior border of the dens and classified > 3 mm as abnormal. The data consisted of 547 normal and 543 abnormal cases, which were randomly split into 1000 training and 90 test datasets. We employed YOLOv5 as the deep learning platform to perform object detection based on the training data with a bounding box label. Our bounding box was placed between the posterior border of the dens and the anterior border of the atlas. We trained our model using the pre-trained YOLOv5l weights with a resolution of 416, and max epoch of 600.

The training was completed in 2.5 hours and 200 epochs, with >0.99 precision and recall. On testing 90 cases, the true positive, true negative, false positive, and false negative counts were 31, 52, 6, and 1, respectively. The accuracy, precision, and recall of the test data were 0.93, 0.84 and 0.97, respectively.

We developed a deep learning model that can aid clinicians in assessing the risk of atlantoaxial instability based on lateral cervical radiographs without immediate input from the radiologist. Our model effectively detects an abnormal ADI on lateral cervical radiographs and may assist physicians in developing further diagnostic or management plans in patients with suspected atlantoaxial instability (AAI).