90. Identification of anterior cervical spinal instrumentation using a smartphone application powered by machine learning

BACKGROUND CONTEXT Anterior cervical discectomy and fusion (ACDF) is a common procedure used in the treatment of the cervical spine. Estimates of revision rate following this procedure range between 10-20%. Identification of existing instrumentation is a critical step in planning revision surgery following ACDF. In many cases, information about the existing instrumentation is unavailable, and identification must be performed visually. This process may be time-consuming and subject to biases. As such, a need exists for fast, objective, automated methods for identifying instrumentation from radiographs. Machine learning algorithms that are known to be adept at image classification may be applied to the problem of ACDF plate identification. Furthermore, due to the ubiquity of smartphone use in clinical practice, an algorithm that could be deployed on a smartphone would be particularly useful. PURPOSE The purpose of this study is to develop and validate a machine learning algorithm for the automated identification of ACDF plates from smartphone images of AP cervical spine radiographs. STUDY DESIGN/SETTING Cross-sectional study. PATIENT SAMPLE A total of 402 smartphone images of deidentified AP cervical spine radiographs containing 15 different types of ACDF plates were gathered from publicly available data sources via internet search. Image selection required implant identities to be confirmable by the authors, so the images could not have severe artifacts. ACDF plates for 1-level, 2-level, and 3-level fusions were included. OUTCOME MEASURES Algorithm performance was evaluated using accuracy, positive predictive value, sensitivity, and f-1 score. Qualitative analysis was performed using saliency heatmaps to evaluate if algorithm decisions were based on reasonable implant features. METHODS The smartphone images of AP cervical spine radiographs were stratified by implant type and divided into a training and a testing dataset. 275 images (∼70%) were used to train and validate a convolution neural network (CNN) for classification of implants from the radiographs. 127 (∼30%) images were held out to test algorithm performance. RESULTS The algorithm performed with an overall accuracy of 94.4% and 85.8% for top-3 and top-1 accuracy respectively. Overall positive predictive value, sensitivity and f1-scores were 0.873, 0.858, and 0.855 respectively. Saliency heatmaps revealed that the algorithm examined unique features of the implants that a human observer might use for implant identification. CONCLUSIONS This algorithm demonstrates strong performance in the classification of ACDF plates from smartphone images and will be deployed as an accessible smartphone application for further evaluation, improvement, and eventual widespread use. FDA DEVICE/DRUG STATUS Aesculap ABC (Approved for this indication), Alphatec Deltaloc (Approved for this indication), Alphatec Deltaloc Reveal (Approved for this indication), Depuy Synthes CSLP (Approved for this indication), Depuy Synthes Skyline (Approved for this indication), Medtronic Atlantis (Approved for this indication), Medtronic Atlantis Vision (Approved for this indication), Medtronic Orion (Approved for this indication), Medtronic Premier (Approved for this indication), Medtronic Zephir (Approved for this indication), Stryker Aline (Approved for this indication), Stryker Aviator (Approved for this indication), Stryker Reflex Hybrid (Approved for this indication), Xtant Spider (Approved for this indication), Zimmer Maxan (Approved for this indication)