In deep learning (DL), a convolutional neural network (CNN) is a model uniquely suited to processing, analyzing, and classifying images. Here, researchers investigated whether a convolutional neural network model trained on lateral cervical spine radiographs can aid in assessing fusion following anterior cervical discectomy and fusion (ACDF). Exploring the use of DL in diagnostic imaging. Following ACDF, they evaluated 187 patients with computed tomography (CT) scans, lateral cervical spine radiographs (both static and motion), and a year’s worth of follow-up.
The CNN-based DL algorithm’s efficacy was measured by computing its accuracy and area under the curve (AUC). CT scans of the neck determined whether the cervical vertebrae had fused. A total of 130 patients (69.5%) were chosen at random as the training set, while the remaining 30.5% (57 patients) were used as the validation set to test the accuracy of the model. Cervical spine x-rays were used to train a convolutional neural network-based DL model. Each patient had their 3 radiographs (in neutral, flexion, and extension) analyzed by the CNN method, and the results were displayed as fusion (0) or nonunion (1).
The final diagnosis of fusion (fusion ≥ 2) or nonunion (fusion ≤ 1) for a patient was reached by weighing the results of all 3 radiographs. The ultimate diagnosis for a patient was either fusion (fusion 2) or nonunion (nonunion ≤ 1) based on the results of all 3 radiographs. The accuracy of the CNN-based DL model was 89.5%, and the area under the curve was 0.889 (95% CI: 0.793-0.984). Pseudarthrosis is difficult to diagnose, but the CNN algorithm for fusion assessment following ACDF trained on lateral cervical radiographs demonstrated a relatively good diagnostic accuracy of 89.5% and is therefore considered a useful aid in the detection process.