gms | German Medical Science

Objectives: The manual measurement of coronal parameters in adolescent idiopathic scoliosis (AIS) is time-consuming and dependent on the physician's experience. Accurate assessment of spinal parameters is crucial for preoperative planning. There is an increasing need for an automated computation of essential radiographic parameters in AIS patients to support clinical practice and basic research.

This study aims to develop and validate a method, which could perform reliable and time efficient measurements of coronal parameters automatically. We hypothesize that the algorithm based on artificial intelligence (AI) will have high agreement with measurements conducted by experienced surgeons for T1-tilt, coronal balance and Cobb angles of proximal thoracic, thoracic and thoracolumbar curves. The analysis of those parameters is globally accepted amongst surgeons as an assessment of spinal deformities. Its automation could support in the accurate and deterministic evaluation of spinal deformities and avoid subjective and error-prone manual measurements.

Methods: A deep learning algorithm was trained on 428 images to detect anatomical structures of interest (cervical, thoracic, lumbar spine and sacrum) in AP radiographs. Based on the spinal curvature extracted from these detections, essential parameters, namely T1-tilt, coronal balance and Cobb angles of proximal thoracic, thoracic and thoracolumbar curves were computed fully automatically.

Two surgeons independently measured 100 preoperative AP spine radiographs. This dataset was completely separate from the training data. To assess intra-rater reliability, one surgeon performed the measurements twice. Mean error, standard deviation as well as inter- and intra-rater reliability between human raters and AI were evaluated using single measure Intraclass Correlation Coefficients (ICC, absolute agreement). ICC>0.75 were considered excellent (Ciccetti, Psychol. Assess. 1994).

Results and conclusion: The comparison between human raters results in excellent ICC values for intra- (range: 0.98-1) and inter-rater (0.85-0.99) reliability. ICC values derived by the comparison between automated measurements and human raters range between 0.77 for proximal thoracic curve and 0.92 for T1-tilt. Exemplarily mean error is smallest for T1-tilt (-0.8°) and largest for thoracic curve (8.4°).

The comparison to digital measurements by human experts demonstrates that the novel AI algorithm delivers highly accurate measurements of essential radiographic coronal spine parameters. It could contribute to efficient, reliable and reproduceable measurements in the treatment of spinal deformities and facilitate the analysis of large datasets (e.g., registry studies) for research purposes.