ESPE Abstracts (2018) 89 FC2.4

Diagnostic Performance of Artificial Neural Network-based TW3 Skeletal Maturity Assessment

Xuelian Zhoua, Junfen Fua, Guanping Donga, Wei Wua, Ke Huanga, Yan Nia, Qiang LINb, Lanxuan Liub, Hao Nib & Can Laia

aThe children’s Hospital of Zhejiang University School of Medicine, Hangzhou, China; bShanghai Yitu Network Technology Co. LTD., Shanghai, China

Purpose: To evaluate the efficacy of supervised Artificial Neural Network (sANN) in bone age assessment and compare the diagnostic performance of sANN-based TW3 skeletal maturity assessment on hand radiographs with that of experienced child endocrinologists.

Materials and methods: This study developed an optimized artificial intelligence TW3 bone age assessment system by using the sANN and rating for the 20 hand bones (13RUS+7carpal) from A to I. 8332 of clinical hand radiographs (the chronological age range from 6 months to 17 years old, male 45.5%, female 54.5%) in our hospital were obtained from Jan 2012 to Dec 2016. Of them 6665 examples were for train set and the rest 1667 ones for validation. 200 independent examples were for the test set, the bone age was firstly assessed by four experienced child endocrinologists according to TW3 rule. Bone age assessment between the artificial intelligence model (BA-AI) and endocrinologists (BA-E) was assessed by comparing the root mean square (RMS) and mean absolute difference (MAD). And we also compared time consumption and stability between model and reviewers.

Results: The mean difference between BA-AI and BA-E was 0.4±0.3 years, with a mean RMS and MAD of 0.54 and 0.48 years. The female bone age difference assessed within 6 months accounted for 92.5, and 98.4% in one year and for males 93.3 and 97.84% respectively. The average time of input hand radiographs to output bone age was 1.5±0.2 s for the model, while 225.6±55.5s for the reviewers. The consistency of TW3 Carpal between the four reviewers is poor than TW3 RUS, the differences between reviewer1 and 2, reviewer 2 and 4 were statistically significant.

Conclusion: The diagnostic performance of sANN-based TW3 skeletal maturity assessment is time saving, the accuracy is similar and the stability is superior to that of experienced endocrinologists.

Keywords: Bone age; Artificial Neural Network; Artificial intelligence; TW3 bone age assessment

Article tools

My recent searches

No recent searches.