Comparison of Three CNN Models Applied in Bone Age Assessment of Pelvic Radiographs of Adolescents.
10.12116/j.issn.1004-5619.2020.05.004
- Author:
Li Qin PENG
1
;
Lei WAN
2
;
Mao Wen WANG
2
;
Zhuo LI
2
;
Peng WANG
3
;
Tai Ang LIU
3
;
Ya Hui WANG
2
;
Hu ZHAO
1
Author Information
1. Guangdong Province Translational Forensic Medicine Engineering Technology Research Center, Faculty of Forensic Medicine, Zhongshan School of Medicine, Sun-Yat Sen University, Guangzhou 510080, China.
2. Shanghai Key Laboratory of Forensic Medicine, Key Laboratory of Forensic Science, Ministry of Justice, Shanghai Forensic Service Platform, Academy of Forensic Science, Shanghai 200063, China.
3. Shanghai Zhenpu Information Technology Co. Ltd., Shanghai 200444, China.
- Publication Type:Journal Article
- Keywords:
forensic anthropology;
age determination by skeleton;
pelvis;
image recognition;
deep learning;
convolutional neural networks;
Han nationality;
adolescents
- MeSH:
Adolescent;
Adult;
Age Determination by Skeleton;
Child;
China;
Female;
Humans;
Male;
Pelvis;
Radiography;
Young Adult
- From:
Journal of Forensic Medicine
2020;36(5):622-630
- CountryChina
- Language:Chinese
-
Abstract:
Objective To compare the performance of three deep-learning models （VGG19, Inception-V3 and Inception-ResNet-V2） in automatic bone age assessment based on pelvic X-ray radiographs. Methods A total of 962 pelvic X ray radiographs taken from adolescents （481 males, 481 females） aged from 11.0 to 21.0 years in five provinces and cities of China were collected, preprocessed and used as objects of study. Eighty percent of these X ray radiographs were divided into training set and validation set with random sampling method and used for model fitting and hyper-parameters adjustment. Twenty percent were used as test sets, to evaluate the ability of model generalization. The performances of the three models were assessed by comparing the root mean square error （RMSE）, mean absolute error （MAE） and Bland-Altman plots between the model estimates and the chronological ages. Results The mean RMSE and MAE between bone age estimates of the VGG19 model and the chronological ages were 1.29 and 1.02 years, respectively. The mean RMSE and MAE between bone age estimates of the Inception-V3 model and the chronological ages were 1.17 and 0.82 years, respectively. The mean RMSE and MAE between bone age estimates of the Inception-ResNet-V2 model and the chronological ages were 1.11 and 0.84 years, respectively. The Bland-Altman plots showed that the mean value of differences between bone age estimates of Inception-ResNet-V2 model and the chronological ages was the lowest. Conclusion In the automatic bone age assessment of adolescent pelvis, the Inception-ResNet-V2 model performs the best while the Inception-V3 model achieves a similar accuracy as VGG19 model.
