Artificial Intelligence-based Models Enabling Robust and Precise Spine Malalignment Auto-analysis in China: a Multicentre, Retrospective Cohort Study
Artificial Intelligence-based Models Enabling Robust and Precise Spine Malalignment Auto-analysis in China: a Multicentre, Retrospective Cohort Study
This retrospective study aimed to enhance and validate a model for diagnosing adolescent idiopathic scoliosis (AIS) across multiple medical centers. The study included 2,763 participants from prestigious hospitals in mainland China and Hong Kong. X-rays were used to develop and validate the model, with data from different hospitals to ensure robustness. Participants aged 10-18 with confirmed AIS were enrolled, and data were deidentified for privacy. The model was optimized using training data and validated internally before being deployed for real-world application. A novel data augmentation technique was used to address data heterogeneity, and a standardized analysis platform, AlignProCARE, was employed for evaluation. X-rays were annotated with vertebra landmarks, and traditional and intensity-based data augmentation methods were applied for image processing. Coronal Cobb angle was used to evaluate spinal alignment, with severity classified as normal-mild, moderate, or severe. The model's performance was statistically assessed for accuracy in predicting Cobb angle and severity grading. Overall, the study aimed to provide a reliable diagnostic tool for AIS analysis in clinical practice, improving efficiency and standardization in diagnosis and treatment.
This retrospective study involved collecting posteroanterior whole-spine X-rays from 2,763 individuals at 5 renowned hospitals in mainland China and 2 hospitals in Hong Kong over a period from January 1, 2012, to April 4, 2021. X-rays from Queen Mary Hospital and Duchess of Kent Children's Hospital at Sandy Bay in Hong Kong, known as the QMH&DKCH cohort, were specifically used for model development. Within this cohort, 86.5% (1686 out of 1950 patients) were randomly chosen as the training set for model development, while the remaining 13.5% (264 patients) formed the internal validation set to ensure the model's performance was independently tested. Additionally, data from five prominent hospitals in mainland China were compiled into external validation datasets to assess the model's efficacy. These hospitals included Peking Union Medical College Hospital, Nanfang Hospital, Jishuitan Hospital, Ruijin Hospital, and Huashan Hospital. All data were deidentified before being utilized for model development and validation. The study enrolled participants aged 10 to 18 years with confirmed presence or absence of adolescent idiopathic scoliosis (AIS). Demographic information such as age, sex, and BMI was extracted from medical records. Exclusion criteria were applied to ensure the study's specificity, excluding individuals with other types of scoliosis, skin diseases that could impact imaging, those unable to stand, and cases where standing imaging was not feasible. The study design, illustrated in Figure 1, consisted of two main phases: model development and real-world application. The model was optimized and validated using the QMH&DKCH cohort data before being deployed on the AlignProCARE platform for clinical application across different centers. In the real-world application stage, five medical centers utilized the AlignProCARE software to upload and analyze patient X-ray images for diagnostic purposes.
Inclusion Criteria:
Exclusion Criteria: