FPDY199
Free Paper (Deformity)
Automated Upper Instrumented Vertebra Tilt Angle Measurement for Adolescent Idiopathic Scoliosis Using Deep Learning Pose Estimation
Tan Shun Herng, Lee Sin Ying, Saturveithan A/L Chandirasegaran, Chiu Chee Kidd, Chan Chris Yin Wei, Kwan Mun Keong
Department of Orthopaedic Surgery, National Orthopaedic Centre of Excellence for Research and Learning (NOCERAL), Faculty of Medicine, Universiti Malaya, Kuala Lumpur, Malaysia
Achieving optimal Upper Instrumented Vertebra (UIV) tilt has been shown to reduce postoperative neck tilt in Lenke 1-2 curves. Optimal UIV tilt angle calculation serves as an intraoperative guide to mitigate neck tilt and shoulder imbalance. However, manual calculation is time-consuming and subject to inter-observer variability. We aimed to develop an artificial intelligence (AI) model to automate UIV tilt angle calculation. This study analysed 60 consecutive AIS patients undergoing posterior spinal fusion. Left and right side-bending radiographs were used to train and validate a YOLOv11-Large pose estimation CNN to detect C7-L4 vertebrae (6 keypoints/vertebrae). The system geometrically calculates vertebral centroids, constructs transpedicular reference lines, and measures UIV tilt angles (T2-T4) following established methodology. Clinical validation included 60 measurements against a senior spine surgeon. Primary outcomes included Intraclass Correlation Coefficient (ICC), Mean Absolute Error (MAE), and Clinical Success Rate (within 3° and 5° thresholds). The AI system achieved robust technical performance with bounding box mAP50 of 99.5% and pose estimation mAP50 of 99.2% on 100-patient detection dataset. In clinical validation, the model demonstrated good reliability with an ICC of 0.761 (95% CI: 0.63- 0.85). The MAE was 1.43°, with systematic bias of
-0.11°(p=0.683). The AI achieved clinical success rates of 91.7% within 3° threshold, and 96.7% within 5° threshold. Component angle agreement was similarly good: RSB (ICC=0.69, MAE=2.05°), LSB (ICC=0.77, MAE=1.98°). Level-specific analysis showed consistent performance: T2 (ICC=0.72), T3 (ICC=0.66), T4 (ICC=0.71).
This study presents the first AI system to automate optimal UIV tilt angle calculation for AIS surgery, achieving good to excellent reliability (ICC=0.76) and clinical accuracy (92% within 3°, 97% within 5°). With mean error of 1.43° and minimal bias, the tool offers a rapid, consistent, and reproducible alternative to manual measurement for routine preoperative planning.