Semiautomated intraoperative measurement of Cobb angle and coronal C7 plumb line using deep learning and computer vision for scoliosis correction: a feasibility study

Abstract

OBJECTIVE Scoliosis is a degenerative disease with a 3D deformity in the alignment of the spinal column. Surgical spinal correction outcomes are heavily dependent on the surgeon’s expertise and use of visual cues because of time requirements, lack of automation, and radiation exposure associated with current intraoperative measurement techniques. In this study, the authors sought to validate a novel, nonradiographic, and semiautomated device that measures spinal alignment intraoperatively using deep learning and computer vision. METHODS To obtain spinal alignment metrics intraoperatively, the surgeon placed 3D-printed markers made of acrylonitrile butadiene styrene (ABS) plastic at designated locations in the surgical field. With the high-definition camera of the device, the surgeon can take an image of the markers in the surgical field. Images are processed through a computer vision model that detects the location of the markers and calculates the Cobb angle and coronal plumb line. The marker detection model was trained on 100 images and tested on 130 images of the ABS markers in various conditions. To verify the Cobb angle calculation, 50 models of angle templates from 0° to 180° in 3.6735° increments were created for testing. To verify the plumb line calculation, 21 models of plumb line measurements from −10 to +10 cm in increments of 1 cm were created for testing. A validation study was performed on a scoliotic cadaver model, and the radiographic calculations for Cobb angle and plumb line were compared with the device’s calculations. RESULTS The area under the curve for the marker detection model was 0.979 for Cobb angle white, 0.791 for Cobb angle black, and 1 for the plumb line model. The average absolute difference between expected and measured Cobb angles on the verification models was 1.726° ± 1.259°, within the clinical acceptable error of 5°. The average absolute difference between the expected and measured plumb lines on the verification models was 0.415 ± 0.255 cm. For the cadaver validation study, the differences between the radiographic and device calculations for the Cobb angle and plumb line were 2.78° and 0.29 cm, respectively. CONCLUSIONS The authors developed and validated a nonradiographic, semiautomated device that utilizes deep learning and computer vision to measure spinal metrics intraoperatively.