Validation of a novel artificial intelligence model (SpinePose) to automatically and accurately predict spinopelvic parameters using scoliosis radiographs in an external cohort.

Objective: SpinePose was developed in 2024 as a novel artificial intelligence (AI) tool to automatically predict spinopelvic parameters with high accuracy and without the need for manual entry. The authors' published results demonstrated excellent performance comparable to a fellowship-trained spine surgeon with more than 15 years of experience. To date, there have not been any studies that have externally validated the performance of AI-based spinopelvic parameter measurement tools on data acquired from other institutions. To assess the generalizability of SpinePose, the authors report its performance on an external set of heterogeneous whole-spine scoliosis radiographs obtained from an outside institution.

Methods: SpinePose was trained/validated on a dataset of 761 sagittal whole-spine scoliosis radiographs from a single institution, with expert-level performance on both whole-spine and lumbosacral radiographs. In this study, the existing SpinePose model was used for inference on a new set of 49 whole-spine radiographs acquired at a tertiary academic hospital located out of state. Externally acquired radiographs represented a diverse set of images, incorporating patients who had undergone instrumentation and those who had not, and a wide variety of fusion constructs including complex deformity patients. Predicted measures included sagittal vertical axis (SVA), pelvic tilt (PT), pelvic incidence (PI), lumbar lordosis (LL), and T1-pelvic angle (T1PA). Predicted parameter values relative to ground-truth manual annotations were calculated to determine the model's accuracy.

Results: Of the 49 images, 35 (71.4%) had instrumentation compared with 51.0% and 57.5% in the original SpinePose training and testing sets, respectively. All 5 parameters in the external dataset were significantly different at baseline compared with the original test set (p < 0.01). SpinePose accurately predicted all 5 spinopelvic parameters without any statistically significant differences: SVA, 50.7 mm vs 52.3 mm (p = 0.85); PT, 27.6° vs 30.5° (p = 0.24); PI, 58.0° vs 61.8° (p = 0.17); LL, 40.4° vs 42.4° (p = 0.77); and T1PA, 24.8° vs 28.0° (p = 0.21) when comparing ground truth annotations with predicted values.

Conclusions: SpinePose was able to accurately predict spinopelvic parameters on an external validation cohort that was generated independently from the images on which the model was trained and validated. This highlights the generalizability of SpinePose to be implemented on novel images from other institutions and geographic settings with high accuracy and minimal preprocessing. The implementation of AI tools more broadly will help standardize our ability to both deliver and provide spine care and assist with surgical treatment and management of spine patients.