Oral Posters: Innovative Technologies
Presented by: V. Lafage - View Audio/Video Presentation (Members Only)
O. Moal(1), B. Moal(2), R. Lafage(1), F. Schwab(1), J. Smith(3), C. Shaffrey(3), A. Jain(4), S. Bess(5), J. Gum(6), G. Mundis(7), C. Ames(8), E. Klineberg(9), V. Lafage(1)
(1) Hospital for Special Surgery, Spine Service, New York, NY, United States
(2) Centre Hospitalier, Universitaire de Bordeaux, Bordeaux, France
(3) University of Virginia Medical Center, Charlottesville, VA, United States
(4) Johns Hopkins University School of Medicine, Department of Orthopaedic Surgery, Baltimore, MD, United States
(5) NYU Langone Medical Center, New York, NY, United States
(6) Norton Leatherman Spine Center, Louisville, KY, United States
(7) Scripps Clinic Torrey Pines, La Jolla, CA, United States
(8) San Francisco Medical Center, University of California, San Francisco, CA, United States
(9) University of California, Davis, Department of Orthopedic Surgery, Sacramento, CA, United States
Introduction: Radiographic analysis and preoperative surgery simulation of ASD spinal reconstructions are important. Several methods exist but are time consuming. Recent progress in machine learning algorithms may offer a novel approach to improve speed and resource utilization. The hypothesis of this study is that a data-driven model for faster and accurate sagittal spine deformity reconstruction based on a few manually identified landmarks is possible.
Method: This study is a retrospective analysis of a prospective database of ASD patients. Sagittal radiographic reconstructions of operated and non-operated ASD patients were analyzed. Each reconstruction was defined by 72 anatomical landmarks: femoral head centers, sacral endplate, and four corners of the lumbar and thoracic vertebrae. After identifying an initial set of 8 landmarks (femoral head centers, sacral plate, L1 and T1 superior endplates), the model was trained to predict vertebral corners between T1 and L1 using PPCA. Improvement of prediction was evaluated by manually identifying additional vertebral corners. RMSE errors in the prediction of vertebral corners were expressed in percentage of the L1 superior endplate length, between true and predicted. The number of additional landmarks to obtain an L1-RMSE below 10% and 5% for 95% of the tested reconstruction was calculated. Average absolute differences in thoracic kyphosis (TK) and thoraco-lumbar (TL) alignment were analyzed.
Results: The model was trained with 1984 reconstructions and tested on 423. After identifying the initial set of landmarks, the RMSE was 16%±7. Four additional landmarks were needed to reach RMSE< 10% for 95% of the tested shapes. Twelve were needed for an RMSE< 5%. Absolute mean errors were 4.8˚ ±3.1 (4 landmarks) and 3.8˚ ±2.5 (12 landmarks) for TK, and 5.3˚±3.4 and 3.8˚ ±2.5 for TL.
Conclusion: Unlike classic models that focus on assessing spino-pelvic parameters, the current work permits prediction of the location of vertebral endplates from T1 to S1 based on the identification of only 12 of the 72 landmarks. This model permits expeditious and accurate full sagittal spine reconstruction.