86. Investigation and validation of deep learning algorithms to identify cervical ossification of posterior longitudinal ligaments on radiography

Abstract

<h3>BACKGROUND CONTEXT</h3> A deep learning algorithm capable of detecting cervical ossification of the posterior longitudinal ligament (OPLL) on radiography has the potential to decrease the incidence of misdiagnosis and OPLL overlook, and facilitate the provision of adequate and timely therapy. Further, it may improve patient safety and decrease radiation exposure by minimizing the need for computer tomography (CT). <h3>PURPOSE</h3> The purpose of present study was to validate the diagnostic yield of our deep learning algorithm created for detecting cervical OPLL on radiography, and to compare its diagnostic accuracy with that of experienced Japanese spine physicians. <h3>STUDY DESIGN/SETTING</h3> Cross-sectional study. <h3>PATIENT SAMPLE</h3> A total of 486 patients (243 patients with OPLL and 243 age and sex matched controls). <h3>OUTCOME MEASURES</h3> The accuracy and the area under curve (AUC) of the receiver operating characteristic curve were calculated. Subsequently, using 50 randomly selected samples, numbers of correct diagnoses were compared between the algorithm and the consensus of four spine physicians. <h3>METHODS</h3> The presence/absence of OPLL was confirmed via CT. The deep learning algorithm, which diagnosed the presence/absence of OPLL on cervical radiography and highlighted areas of ossification in positive cases, was constructed using a convolution neural network model called EfficientNetB2 with radiography and CT scan data. <h3>RESULTS</h3> The diagnostic accuracy of our deep learning algorithm was 0.88; and the AUC determined was 0.94. Results could have been affected by institutions where radiographic images were taken, and the type and level of the OPLL. Our deep learning algorithm had significantly more correct diagnoses than did the spine physician consensus (47/50 vs 39/50, respectively; p = 0.041, Fisher's exact test). <h3>CONCLUSIONS</h3> The accuracy of our deep learning algorithm in the diagnosis of cervical OPLL on radiography was significantly higher than that of the experienced spinal physician consensus. The algorithm may improve OPLL diagnostic accuracy by using different diagnostic criteria than physicians, and may help both physicians and patients by facilitating the provision of adequate therapy with minimal radiation exposure. <h3>FDA DEVICE/DRUG STATUS</h3> This abstract does not discuss or include any applicable devices or drugs.