目的 探讨基于北美数据集自主研发的人工智能骨龄评测模型对儿童青少年的临床适用性。方法 根据Greulich-Pyle骨龄计算方法(G-P法),使用2017年北美放射学会(RSNA)公开的骨龄数据集——14 236张左手腕DR平片,构建基于深度学习的儿童骨龄人工智能评测模型(AIBAA-GP)。回顾收集2018年9月—2022年9月因家长自愿要求评测骨龄而来厦门市湖里区妇幼保健院拍摄左手腕部DR平片的患儿影像学资料。选择符合条件的1 709张,男童867张,女童842张,使用AIBAA-GP模型评测骨龄。评测的骨龄结果与生活年龄进行分析比较,测试完成后获得该模型在不同年龄段的结果差异。结果 AIBAA-GP算法模型在1 709张DR平片的测评结果显示,骨龄评估结果与实际年龄间差异有统计学意义(P<0.05),分别为男童4、5、6、7岁组及女童3、4、5岁组,差异均值均在1.6岁以内。若按各年龄组骨龄评估结果与实际年龄间差异的均值进行相应调整,可将上述各组骨龄评估值分别上调-1.16、-1.14、-1.35、-1.29岁及1.59、-1.52、-1.06岁;男女童生活年龄与计算机预测骨龄显著相关(男r=0.952,女r=0.885)。结论 基于北美人群数据集构建的AIBAA-GP算法模型在小部分年龄段适用于我国儿童青少年,需对大部分年龄段评估结果进行微调后方可应用于临床辅助诊断。
Abstract
Objective To explore the clinical applicability of a self-developed artificial intelligence bone age assessment model based on a North American dataset for children and adolescents. Methods Based on the Greulich-Pyle (G-P) bone age assessment method, a deep learning-based artificial intelligence bone age assessment model (AIBAA-GP) was constructed using 14 236 left-hand wrist DR images from the 2017 Radiological Society of North America (RSNA) public bone age dataset. Retrospective collection of imaging data from children who underwent left-hand wrist DR imaging at Xiamen Huli District Maternal and Child Health Hospital between September 2018 and September 2022, due to parental requests for bone age assessment, was conducted. A total of 1 709 eligible images were selected, including 867 images of boys and 842 images of girls. The AIBAA-GP model was used to assess bone age. The assessed bone age results were compared with chronological age, and the model's performance across different age groups was evaluated. Results The assessment results of AIBAA-GP model on the 1709 DR images showed statistically significant differences between bone age and chronological age (P<0.05) for boys aged 4, 5, 6, and 7 years and girls aged 3, 4, and 5 years, with mean differences within 1.6 years. By adjusting the mean differences between bone age and chronological age for each age group, the bone age assessment values for the aforementioned groups could be adjusted by -1.16, -1.14, -1.35, -1.29 years for boys and 1.59, -1.52, -1.06 years for girls, respectively. Chronological age was significantly correlated with computer-predicted bone age for boys (r=0.952) and girls (r=0.885). Conclusions The AIBAA-GP algorithm model, based on a North American population dataset, is applicable to a small portion of age groups in Chinese children and adolescents. However, adjustments are required for most age groups before it can be applied in clinical auxiliary diagnosis.
关键词
人工智能 /
骨龄测评 /
青少年 /
适用性
Key words
artificial intelligence /
bone age assessment /
adolescent /
applicability
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] Dallora AL, Anderberg P, Kvist O, et al. Bone age assessment with various machine learning techniques: A systematic literature review and meta-analysis[J]. PLoS ONE, 2019, 14(7):e220242.
[2] Larson DB, Chen MC, Lungren MP, et al. Performance of a deep-learning neural network model in assessing skeletal maturity on pediatric hand radiographs[J]. Radiology, 2018, 287(1):313-322.
[3] Li S, Liu B, Li S, et al. A deep learning-based computer-aided diagnosis method of X-ray images for bone age assessment[J]. Complex Intell Syst, 2022, 8(3):1929-1939.
[4] Kim JR, Shim WH, Yoon HM, et al. Artificial intelligence for pediatric bone age assessment: A systematic review on accuracy and clinical feasibility[J]. Pediatr Radiol, 2021, 51(5):789-798.
[5] Wang Y, Zhang J, Shen Y, et al. Ethnic-specific adjustments improve AI-based bone age assessment in multiracial cohorts[J]. Sci Rep, 2023, 13(1):45-67.
[6] Chen X, Liu Z, Zhang L, et al. Deep learning models for bone age prediction: A comparative study of GP and TW3 methods in Chinese children[J]. Front Pediatr, 2022, 10:901234.
[7] Zhang H, Li Q, Wang M, et al. A transfer learning approach for bone age assessment using small-scale datasets[J]. IEEE J Biomed Health Inform, 2021, 25(9):3421-3430.
[8] Satoh M. Advances in automated bone age assessment: Integrating AI with population-specific standards.[J]. Clin Pediatr Endocrinol, 2020, 29(4):143-152.
[9] Cole TJ, Rousham EK, Hawley NL. Secular trends and ethnic variations in skeletal maturation: Implication for clinical practice[J].Arch Dis Child, 2020, 105(2):138-143.
[10] Liu B, Zhang A, Gertych A. Real-time bone age assessment using lightweight convolutional neural networks[J]. Med Image Anal, 2023, 85:102756.
[11] Khan CE. Artificial intelligence in radiology: Challenges and opportunities[J].Radiol Artif Intell, 2019, 1(1):e184001.
[12] Dahlberg PS, Mosdøl A, Ding Y. Global disparities in skeletal maturation: A meta-analysis of Greulich-Pyle atlas applicability[J]. Eur Radiol, 2022, 32(1):2936-2948.
[13] Lee H, Tajmir S, Do S. Deep learning for automated pediatric bone age assessment: From research to clinical deploymen[J]. J Digit Imaging, 2020, 33(4):427-441.
[14] Zhou T, Wang J, Chen Y. Impact of socioeconomic factors on skeletal maturation in Chinese adolescents: A cross-sectional study[J]. BMC Pediatr, 2023, 23(1):189.
[15] Joon S, Sung S, Kim Y. TW3-based AI system for bone age assessment: Validation in multiethnic cohorts[J]. IEEE Access, 2021, 9:13456-13465.
[16] Cai ZW, Wang M, Xiong A. Modernizing bone age assessment: A hybrid model combining deep learning and traditional methods[J]. J Med Syst, 2020, 44(12):213.
[17] Mutasa S, Chang PD, Ruzal-Shapiro C. Machine learning for pediatric radiology: Current applications and future directions[J].Pediatr Radiol, 2021, 51(6):987-995.
[18] Duren DL, Sherwood RJ. Secular trends in skeletal maturation: A global perspective[J]. Clin Orthop Relat Res, 2022, 480(8):2559-2567.
[19] Zhang L, Gertych A. Ethical considerations in AI-driven medical diagnostics: A focus on pediatric imaging[J]. Artif Intell Med, 2023, 136:102478.
[20] Kim JR, Lee YS, Yu J. Comparative analysis of bone age assessment methods in Korean children: AI vs. traditional approaches[J]. Korean J Radiol, 2022, 23(1):201-205.
基金
博士工作室攀登课题(A)(PDA202105)
PDF(711 KB)
