Deep Learning-Based Quantitative Measurement Study of Spino-Pelvic Parameters in Adolescent Idiopathic Scoliosis Patients

Objective To develop a deep learning (DL) model for the automated measurement of spino-pelvic parameters on whole-spine digital radiography (DR) images of adolescent idiopathic scoliosis (AIS) patients and to evaluate its performance and generalizability. Methods A total of 1348 whole-spine frontal/lateral DR images of AIS patients were collected to train the DL model for predicting fourteen spino-pelvic parameters. The reliability of manual annotation and the model's performance in detecting key points were assessed using Percentage of Correct Keypoints (PCK) values. Differences between the model-predicted parameter values and manual measurements, used as the reference standard, were analyzed using paired t-tests. Further performance evaluations included the mean absolute error (MAE), Pearson correlation coefficient (r), intra-class correlation coefficient (ICC), Bland-Altman plots. Results The DL model demonstrated the ability to automatically detect vertebral bodies and key points. The PCK for detecting vertebral body key points in whole-spine DR images ranged from 82.2–95.3% within a 3-mm threshold. Additionally, the PCK for the left and right femoral heads in whole-spine lateral DR images was 77.9% and 62.3%, respectively. The spino-pelvic parameter values measured by the DL model exhibited a high correlation and agreement with the reference standard (ICC: 0.9–1.0, r: 0.8–1.0, MAE: 1.0–3.7).The Whole frontal model used VF-Net outperformed other networks (one stage HRNet、one stage SCNet、two stage HRNet-HRNet、two stage SCNet-SCNet) in predicting landmarks within a distance threshold of 2.5 to 5 mm;The Whole lateral model used two stage HRNet-HRNet outperformed other networks (VF-Net,one stage HRNet、one stage SCNet、two stage SCNet-SCNet) in predicting landmarks within a distance threshold of 1 to 5 mm. Conclusions The DL model developed in this study can automatically measure spino-pelvic parameters with performance comparable to that of radiologists. This model is expected to provide an automated measurement tool for clinical practice, thereby improving efficiency in diagnosing, monitoring, and treating AIS patients.