Performance of a deep learning-based algorithm for automated measurements of Cobb angles on preoperative spine radiographs in adolescent idiopathic scoliosis.

Authors

Abstract

Accurate Cobb angle measurement is essential in adolescent idiopathic scoliosis (AIS), but evidence on artificial intelligence (AI) performance in pediatric patients, especially with severe curves, is limited. This study evaluated the accuracy of a commercially available deep learning software in measuring Cobb angles in surgical cases of AIS and compared its performance with that of radiology residents. A total of 151 preoperative anteroposterior whole spine X-rays were analyzed. The ground truth was established by consensus between a pediatric radiologist and a pediatric orthopedic surgeon. The mean absolute error (MAE) of the AI and six radiology residents was calculated. Wilcoxon signed-rank test and intraclass correlation coefficients (ICC) were computed. The dataset included 151 angles categorized as moderate (13; 20-39°), severe (74; 40-59°), and extreme (64; ≥60°). Overall, the AI's MAE was 4.57° [95%CI 3.85°, 5.29°], significantly higher than that of the residents' (P=0.0017). Scoliosis severity significantly affected MAE in both groups. For extreme scoliosis, the MAE of the AI (6.53° [95%CI 5.06°, 7.86°]) was significantly higher than that of the residents (3.61° [95%CI, 3.17°, 3.97°]) (P<0.001). No significant difference in MAE was observed between patients with moderate and severe scoliosis (P=0.86). The agreement between the ICC between the AI and the ground truth was 0.89 [95%CI 0.69, 0.95]. The Cobb angle measurements obtained with this software agree with those of experts for moderate and severe scoliosis with a clinically acceptable average error, without significant difference compared to radiology residents. AI accuracy significantly decreases in patients with extreme scoliosis, highlighting the need for radiologist oversight.

Tags

Topics