Related Experiment Video
Updated: Jan 15, 2026
Longitudinal Micro-Computed Tomography Image Analysis for User-Defined Region of Interest in Critical-Sized Bone Defects
Published on: June 24, 2025
Radiomics-powered radiographic image analysis for enhanced mechanical complications prediction and surgical planning
Ferran Pellisé1, Sleiman Haddad1, Susana Núñez Pereira1
1Spine Surgery Unit, Vall d'Hebron University Hospital, Barcelona, Spain.
Background Context:
Radiomics, a technique employing machine learning (ML) to extract quantitative features from processed radiographic images, holds promise for improving clinical prediction models. It offers the potential to comprehensively characterize spinal shape and alignment. We hypothesized that processed image (PrIm) algorithms outperform traditional radiographic measurements (TRM) and scores in predicting postoperative mechanical complications (MC) in adult spinal deformity (ASD).
Purpose:
The aim was to compare the performance of PrIm algorithms to TRM-GAP score in the prediction of MC in ASD patients.
Study Design/Setting:
An AI-leveraged retrospective analysis was conducted using data from a prospective international multicenter database dedicated to ASD.
Patient Sample:
The study focused on ASD patients aged 18 or older who were surgically treated and who had a minimum follow-up period of two years, with complete preoperative, 6-week and 2-year follow-up.
Outcome Measures:
Major mechanical complications such as rod fractures, pseudarthrosis, or junctional kyphosis or failure.
Methods:
Processed full-spine standing radiographic images were analyzed using an automatic vertebral centroid generation algorithm to map posteroanterior (PA) and lateral spinal shape. Distances and angles between each vertebra and the pelvic centroid were automatically obtained. Machine learning models were constructed using Catboost, combining nonradiographic variables (Non-R: demographic, PROMS, surgical), TRM + GAP score, and PrIm features. AUC-ROC, sensitivity, specificity, and Brier score (0= perfectly calibrated / 1=poor) were used to evaluate prediction accuracy. SHapley Additive exPlanations (SHAP) values were employed to assess variable contributions and address overfitting/noise.
Results:
A total of 690 patients (81% female, 52±19 years, 9.7±3.9 levels, 18.6% 3CO, 43.5% pelvic fixation, 37.5% MC) were analyzed. The Non-R + PrIm model outperformed the present "Gold Standard" model (Non-R + TRM-GAP): AUC-ROC 0.75 vs 0.71 (p=.009), accuracy 0.72 vs 0.62 (p<.001), specificity 0.79 vs 0.60 (p<.001), sensitivity 0.52 vs 0.70 (p<.001), and Brier score 0.17 vs 0.22 (p<.001). Adding TRM and GAP score to Non-R + PrIm model did not improve model estimates. SHAP adjusted models summed 35 variables and revealed PrIm's superior predictive importance, contributing 65.7% to the model compared to Non-R (Surgical factors 16.1%, PROMS 11.3% and demographics 6.9%). Personalized SHAP decision plots identified the most critical vertebral centroids associated with MC risk both globally and individually.
Conclusion:
Radiomics powered by full-spine processed radiographic images enable the most accurate predictive models for MC in ASD. This novel approach offers clinicians a powerful and time-efficient tool for personalized surgical planning, ultimately enhancing ASD surgical outcomes.
