Artificial intelligence can support spine alignment assessment in adolescent idiopathic scoliosis, but performance often falls when radiographs come from different hospitals with different imaging protocols and post-processing methods. A recent study published in the Journal of Medical Internet Research reported a real-time pixel intensity–based transformation designed to reduce those differences while preserving anatomical integrity. The approach was integrated into SpineHRNet+ and evaluated on 3899 full-spine radiographs from 7 hospitals, using Hong Kong data for training and internal validation and mainland datasets for external validation. The results showed more consistent Cobb angle prediction and severity grading across heterogeneous imaging environments.
A Workflow Built for Heterogeneous Imaging
The study design reflected the practical reality of multicentre imaging. Radiographs were gathered across more than a decade, and the participating hospitals used different imaging systems and workflows, including EOS, Philips, GE and several devices at one site. The variation was important because it captured the kind of heterogeneity that often limits real-world deployment. To support training and evaluation, all radiographs were annotated using the AlignProCARE system with 72 vertebral endplate landmarks from C7 to L5. Junior clinicians produced the initial annotations in the training centres, while senior spine specialists verified them. External annotations were also cross-checked, and a 10% sample from each centre was reannotated and reviewed by an independent expert panel to confirm intercentre consistency. The coronal Cobb angle was used as the main measure of spinal alignment, with severity grouped into normal-mild, moderate and severe ranges, and curve type defined by the position of the apex.
The enhanced workflow had 2 stages: model development and multicentre validation. During development, the transformation method was applied to the training set to create simulated radiographs that reflected the pixel-level characteristics of external centres. These transformed images were then mixed with the original images during training and fine-tuning. In the validation stage, each centre uploaded radiographs through locally installed AlignProCARE software or applications, the enhanced model processed the images remotely, and results were returned for clinical analysis. This structure made the model relevant not only as a research exercise but as a system designed for deployment across institutions with different imaging conditions.
Must Read: AI Shows Strong Accuracy for Detecting Hepatic Steatosis
Pixel-Level Harmonisation Without Anatomical Distortion
The central technical contribution was an intensity distribution–based transformation method designed to reduce differences in image appearance without changing clinically important anatomy. Instead of relying on paired training data or more complex generative approaches, the system analysed statistical profiles of pixel intensity distributions for each institution and then aligned images to a reference standard through histogram matching and optimisation. The mapping was monotonic, which allowed the method to preserve anatomical structure while reducing discrepancies in brightness and contrast. The study describes the process as unsupervised, real-time and computationally efficient, making it suitable for routine use rather than only offline experimentation.
The transformation was also designed to maintain diversity during training. Reference images from external centres were randomly sampled, and each training image could be transformed multiple times using different reference distributions. Those augmented samples were then mixed with original images without weighting, which helped avoid overfitting to any single external style. The study reports that this approach reduced the initial differences in histogram spread and brightness between internal and external datasets from 9.6×10-4 and 8.71 to 0.4×10-4 and 0.47. The method therefore addressed a concrete source of model instability while remaining lightweight enough for rapid integration into clinical workflows.
More Stable Performance Across Centres
The performance findings were strongest in the external datasets, where robustness mattered most. In internal validation, the enhanced model produced modest reductions in mean absolute error for both thoracic and thoracolumbar or lumbar curves. In the external datasets, the gains were larger. Mean absolute error fell from 5.38° to 4.03° for thoracic curves and from 5.21° to 3.75° for thoracolumbar or lumbar curves. The study also reports statistically significant improvements across all external cohorts except the thoracic group at PUMCH. These results indicate that the transformation method strengthened the model where heterogeneity had the greatest impact.
Other performance measures pointed in the same direction. The enhanced model maintained an R² above 0.90 in all cohorts and reached 0.93 in the internal dataset. Bland-Altman analysis showed narrower limits of agreement across all datasets, while the spread of prediction errors also decreased. In disease severity grading, the enhanced model achieved macro-average sensitivity of 89.97% and negative predictive value of 94.49% in the internal dataset. Across the 5 external datasets, sensitivity reached 90.18% and negative predictive value 93.16%, improving on the original model in both measures. Curve type classification improved slightly in internal validation, although performance varied across external centres. Taken together, the results show that the enhanced model was not simply more accurate in a narrow sense, but more stable across institutions with different imaging characteristics.
The study showed that integrating a real-time data transformation method into SpineHRNet+ improves performance in multicentre adolescent idiopathic scoliosis assessment by reducing variation in radiographic data across institutions. The approach improved consistency and accuracy across heterogeneous imaging settings while preserving anatomical integrity, which supports reliable use in clinical practice. These results reinforce the importance of addressing data variability when AI models are deployed across different healthcare environments and position data harmonisation as a practical requirement for scalable clinical implementation.
Source: Journal of Medical Internet Research
Image Credit: iStock
References:
Chen G, Meng N, Zhuang Y et al. (2026) Scalable and Robust Artificial Intelligence for Spine Alignment Assessment: Multicenter Study Enabled by Real-Time Data Transformation.
J Med Internet Res;28:e78396.
