Computer vision successfully identified posterior thoracolumbar instrumentation systems
Click Here to Manage Email Alerts
A computer vision detector with a support vector machine classifier more accurately and efficiently identified posterior thoracolumbar hardware at five-level classification compared with surgeons and manufacturer representatives.
Researchers used KAZE feature detection, a multiscale 2D feature detection and description algorithm, to construct a computer vision support vector machine classifier that analyzed 355 lateral, 379 anteroposterior (AP) and 338 fused radiographs from patients undergoing posterior thoracolumbar pedicle screw implantation.
Researchers compared the performance of the computer vision model with the current practice of surgeons and manufacturer representatives identifying hardware by visual inspection. According to the study, the five pedicle screw implants were the Creo (Globus Medical), Solera (Medtronic), Reline (NuVasive), Xia (Stryker) and Expedium (DePuy Synthes).
Overall, computer vision successfully classified the five hardware systems on 100 test images with 79% accuracy in 14 seconds. In comparison, a team of two surgeons and three manufacturer representatives successfully classified the five hardware systems on 100 test images with 44% accuracy in 20 minutes.
Researchers noted the multilevel five-way classification accuracy rates were 64.27% for lateral images, 60.95% for AP images and 65.9% for fused images. However, when Globus Medical and Medtronic, the institution’s two most common manufacturers, were binarily classified, the accuracy rates were 93.15% for lateral images, 88.98% for AP images and 91.08% for fused images. Researchers also noted speeded up robust features, maximally stable extremal regions and minimum eigenvalue feature detectors were tested; however, these had difficulty identifying differences between screws.
“This machine learning approach demonstrated good accuracy among five manufacturers in three radiographic views and showed superiority compared with expert human review in both accuracy and efficiency,” the researchers wrote in the study. “The relative computational simplicity of our model may help facilitate future studies by prospectively analyzing the efficacy of machine learning in clinical settings,” they concluded.
Reference:
Alcantarilla PF, et al. Computer Vision – ECCV. 2012;doi:10.1007/978-3-642-33783-3_16.
Collapse
Click Here to Manage Email Alerts