A research team led by Professor Dong-Ho Kang of the Department of Orthopedic Surgery, Spine Center, has developed an AI-powered automated cervical spine alignment measurement system and validated its superiority through multinational external verification. The findings were published in npj Digital Medicine (IF= 15.2), a Nature Partner Journal, garnering significant attention from the academic community.
A longstanding challenge in cervical spine radiography has been the frequent obstruction of the C7 vertebral body — the lowermost cervical vertebra — by the patient's shoulders, which impedes accurate measurement.
To directly address this limitation, the research team, under Professor Dong-Ho Kang’s leadership, independently developed a multi-stage hierarchical deep learning pipeline that deliberately incorporated C7-obstructed images into the training dataset, thereby overcoming this technical barrier.
The model was designed with a three-stage architecture consisting of: ▲initial prediction, ▲C7 region detection, and ▲high-resolution precision correction, trained on a multinational dataset of 5,604 images. Even under rigorous external validation conditions where C7 obstruction was present in 82% of cases, the model demonstrated high accuracy and reproducibility with measurements closely aligned with those of specialist physicians.
Notably, the level of agreement between the AI system and individual spine specialists exceeded the inter-rater agreement observed between two spine specialists — demonstrating that AI can provide more consistent and reliable measurements than human examiners. Furthermore, in cases where a single-model approach yielded large errors exceeding 10°, the hierarchical model was shown to automatically correct these to within 0.22°, confirming its robust error recovery capability.
"If the AI system developed through this research is implemented in clinical settings," said Professor Dong-Ho Kang, "it is expected to significantly reduce cervical spine image interpretation time and minimize inter-observer variability, enabling more accurate and standardized spinal care."
The research team has developed an AI software prototype based on this model and is currently conducting a proof-of-concept pilot in a real clinical environment. Plans are underway to pursue further validation with the goal of commercializing the technology as a certified medical AI product.