Intraoperative full-spine imaging with C-arm using a deep learning-based automatic stitching model: development and clinical validation.

PURPOSE: To develop and clinically validate an automated deep learning–based stitching model for intraoperative generation of full-length spinal radiographs from segmented C-arm fluoroscopy during scoliosis surgery. METHODS: A deep learning model was developed to detect pedicle screws and align sequential intraoperative C-arm images to reconstruct a full-length spine view. For clinical validation, we retrospectively analyzed 43 scoliosis patients (2018–2023) with adequate intraoperative segmented sequences. Stitched intraoperative images (Method A) were compared with postoperative full-length standing digital radiographs as the reference standard (Method B). Five spine surgeons independently measured coronal Cobb angles on both image sets. Interobserver reliability was assessed using intraclass correlation coefficients (ICC). Agreement between methods was evaluated using ICC, Bland–Altman analysis, and absolute paired differences. RESULTS: The model generated stitched full-length images within 5 s. The mean absolute Cobb angle difference between Methods A and B was 1.95°(SD 2.81°), with a median of 0.20°(IQR 0.00–2.80°) and a range of 0.00–9.80°. Absolute differences were ≤ 1°, ≤ 3°, and ≤ 5° in 76.74%, 88.37%, and 97.67% of paired measurements. Interobserver reliability was excellent (ICC 0.999 for Method A; 0.998 for Method B). Between-method agreement was high (ICC 0.991) with minimal bias (mean difference − 0.13°, 95% limits of agreement − 4.20° to 3.93°). CONCLUSION: Automated stitching of routine intraoperative C-arm images can rapidly produce full-length spinal radiographs with excellent Cobb angle agreement versus postoperative DR, supporting intraoperative alignment assessment.