基于人工智能的骨龄评估工具与Greulich-Pyle及Tanner-Whitehouse第2版方法在儿童队列中的概念验证比较
本研究为概念验证研究,纳入203例因疑似内分泌或生长相关疾病行手及腕部X线摄影的儿童(平均年龄9.0岁,范围2.0-17.0岁),排除技术不合格图像后157例分析AI与GP法,35例子集由儿童内分泌科医生用TW2法评估。以实际年龄为参考,AI的MAE为1.38年,与放射科医生GP法的1.30年相当,优于TW2法的2.86年;AI偏倚极小(-0.05年),TW2系统性低估骨龄(偏倚-2.63年)。结论认为BoneView与放射科医生GP法准确性相当且优于TW2,AI系统可能提高儿童骨龄评估的一致性,但需仔细验证,尤其在种族多样性人群中。
要点
- AI骨龄评估MAE 1.38年,与放射科医生GP法1.30年相当,优于TW2法2.86年。
- AI偏倚-0.05年,TW2系统性低估骨龄偏倚-2.63年。
- AI与GP和实际年龄强相关(r=0.857和0.894),TW2相关性不强(r=0.490)。
核心结论
- BoneView与放射科医生GP法准确性相当,且在本队列中优于TW2评估;AI系统可能提高儿童骨龄评估的一致性,但需仔细验证,尤其在种族多样性人群中。
研究设计与样本
- 概念验证研究,纳入203例因疑似内分泌或生长相关疾病行手及腕部X线摄影的儿童(平均年龄9.0岁,范围2.0-17.0岁),排除技术不合格图像后157例用AI和GP法分析,其中35例子集由儿童内分泌科医生用TW2法评估;以实际年龄为参考。
诊断标准与关键数值
- 以实际年龄为参考,AI的MAE为1.38年、RMSE为1.75年、偏倚-0.05年、r=0.857;放射科医生GP法的MAE为1.30年、RMSE为1.80年、r=0.894;TW2法的MAE为2.86年、RMSE为3.88年、偏倚-2.63年、r=0.490。
临床应用与阅片要点
- 作者认为AI系统可能提高儿童骨龄评估的一致性,但需仔细验证,尤其在种族多样性人群中。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
BACKGROUND: Bone age assessment is essential in evaluating pediatric growth disorders. Artificial intelligence (AI) systems offer potential improvements in accuracy and reproducibility compared to traditional methods.
OBJECTIVE: To compare the performance of a commercially available artificial intelligence-based software (BoneView BoneAge, Gleamer, Paris, France) against two human-assessed methods-the Greulich-Pyle (GP) atlas and Tanner-Whitehouse version 2 (TW2)-in a pediatric population.
MATERIALS AND METHODS: This proof-of-concept study included 203 pediatric patients (mean age, 9.0 years; range, 2.0-17.0 years) who underwent hand and wrist radiographs for suspected endocrine or growth-related conditions. After excluding technically inadequate images, 157 cases were analyzed using AI and GP-assessed methods. A subset of 35 patients was also evaluated using the TW2 method by a pediatric endocrinologist. Performance was measured using mean absolute error (MAE), root mean square error (RMSE), bias, and Pearson's correlation coefficient, using chronological age as reference.
RESULTS: The AI model achieved a MAE of 1.38 years, comparable to the radiologist's GP-based estimate (MAE, 1.30 years), and superior to TW2 (MAE, 2.86 years). RMSE values were 1.75 years, 1.80 years, and 3.88 years, respectively. AI showed minimal bias (-0.05 years), while TW2-based assessments systematically underestimated bone age (bias, -2.63 years). Strong correlations with chronological age were observed for AI (r=0.857) and GP (r=0.894), but not for TW2 (r=0.490).
CONCLUSION: BoneView demonstrated comparable accuracy to radiologist-assessed GP method and outperformed TW2 assessments in this cohort. AI-based systems may enhance consistency in pediatric bone age estimation but require careful validation, especially in ethnically diverse populations.
引用
- Marinelli L, Lo Mastro A, Grassi F, et al. Proof-of-concept comparison of an artificial intelligence-based bone age assessment tool with Greulich-Pyle and Tanner-Whitehouse version 2 methods in a pediatric cohort. Pediatr Radiol. 2025. PMID: 40996527. doi:10.1007/s00247-025-06405-0
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。