Accuracy of general-purpose large language models (LLMS) in interpreting adolescent idiopathic scoliosis radiographs: a comparative study with experienced spine surgeons.

PURPOSE: This study evaluates the accuracy of general-purpose, vision-capable large language models (LLMs) in interpreting adolescent idiopathic scoliosis (AIS) radiographs and compares their outputs with an expert surgeon consensus reference standard.

METHODS

This retrospective comparative diagnostic accuracy study analyzed 128 anonymized AIS radiographic case sets from 128 different patients, including standing coronal and sagittal radiographs and lateral bending films. ChatGPT GPT-5.2 Instant, Gemini 3 Flash, and Grok 4.1 were evaluated through their official web interfaces in January 2026 using the original standardized prompt based on the Lenke classification system. Model inputs were de-identified raw radiographs without measurement lines, Cobb values, vertebral labels, or Surgimap overlays. Three spine surgeons with more than 10 years of deformity surgery experience independently assessed all cases using Surgimap. Surgeon consensus was the reference standard. Cobb angle (CA) accuracy was assessed using mean absolute error and agreement within ± 5°; categorical variables were assessed by exact agreement with 95% confidence intervals.

RESULTS

ChatGPT achieved 94.5% agreement within ± 5° of surgeon consensus, with a mean absolute CA error of 2.4°. Gemini achieved 92.2% agreement with a mean absolute error of 2.7°, and Grok achieved 90.6% agreement with a mean absolute error of 3.1°. Exact agreement was substantially lower for Lenke classification, lumbar and sagittal modifiers, neutral vertebra, and stable vertebra identification. Inter-rater reliability among surgeons was high for CA and categorical parameters. No formal inferential ranking between LLMs was performed.

CONCLUSION

The tested general-purpose LLM versions showed limited reliability for comprehensive AIS radiographic interpretation. Their performance was substantially weaker for Lenke classification, modifiers, and fusion-level-related landmarks than for CA estimation. The findings do not support independent use of these systems for scoliosis measurement, education, or surgical planning. At present, these tools should be regarded as supervised documentation or teaching aids rather than measurement or decision-making systems.