MD Candidate University of Michigan Medical School Ann Arbor, Michigan, United States
Introduction: Over 60% of aging adults suffer from an adult spinal deformity (ASD). Spinopelvic parameters reflecting global sagittal spine imbalance have an established significant correlation with symptom severity. The automated calculation of spinopelvic parameters may serve as a key element of using computer-assistance for patient selection and prediction of surgical outcomes. We developed an artificial intelligence program using deep learning to automatically measure spinopelvic parameters on spinal x-rays.
Methods: We analyzed an image dataset of 311 sagittal scoliosis spinal x-rays. In each image, a set of 8 keypoints were manually labelled at C7, T1, L1, S1, and both femoral heads. Our program trained an open-source neural network (Pytorch Keypoint RCNN) on a subset of these images and predicted 8 keypoint coordinates on a validation set. Spinopelvic parameters were calculated post-hoc from the predicted keypoints. An average Manhattan loss (predicted – ground truth) was calculated at each spinopelvic parameter to measure model accuracy.
Results: Spinopelvic parameters were automatically calculated without human input on 40 patients who were excluded from the training set. The automatic calculations were compared to manual measurements. The following average spinopelvic loss values were determined among the 10-90th percentile range: Sagittal Vertical Axis- [6.81mm ± 0.84mm]; Pelvic tilt- [3.02° ± 0.34°]; Sacral slope- [4.84° ± 0.40°]; Pelvic incidence- [6.69° ± 0.53°]; and Lumbar lordosis- [8.85° ± 0.80°].
Conclusion : Our model shows promising accuracy in automatically measuring spinopelvic parameters and suggests future opportunities for clinical use in predictive analytics. This was accomplished using a small training dataset with relatively few labelled anatomic keypoints. Ongoing modifications including enhanced image processing as well as refined image exclusion criteria for patient positioning and landmark visibility will be explored for improved performance.
