Artificial intelligence-based bone age assessment using deep learning of characteristic regions in digital hand radiograph
10.3760/cma.j.issn.1005-1201.2019.10.020
- VernacularTitle: 基于手腕部影像传统关注特征区域深度学习的人工智能骨龄评估
- Author:
Ying WEN
1
;
Xuhua REN
2
;
Xiujun YANG
1
;
Lihong LI
1
;
Jun LAN
3
;
Tingting LI
1
;
Qian WANG
2
;
Lili SHI
1
Author Information
1. Department of Radiology, Shanghai Children′s Hospital, Shanghai Jiao Tong University, Shanghai 200062, China
2. School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200030, China
3. Winning Health Technology Group Co., Ltd., Shanghai 200072, China
- Publication Type:Journal Article
- Keywords:
Age determination by skeleton;
Radiography;
Diagnosis, computer-assisted;
Artificial intelligence
- From:
Chinese Journal of Radiology
2019;53(10):895-899
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To detect the feasibility and efficiency of bone age(BA) artificial intelligence(AI) estimation based on deep learning features from traditional regions of interest(ROI) in hand digital radiographs(DR).
Methods:BA dataset of left hand DR with 11 858 subjects aged from 0 to 18 years in Children′s Hospital of Shanghai were split to training(80.0%) and validation (20.0%) set in this study. An improved regression convolutional neural networks and extreme gradient boosting decision tree method were utilized for the BA analysis based on traditional ROIs in the images. Another set of BA data with 1 229 subjects also in the hospital was adopted for test. Mean average precision(mAP) and mean absolute error(MAE) were used to assess model accuracy of detection and BA prediction, respectively.
Results:The mAP of ROIs detection of the model was 0.91,and MAE of all male and female subjects was 0.461 and 0.431 years respectively in validation and test sets. The difference less than 1 year in test accounted for 90.07% between BA assessment of the model and of the peadiatric radiologists, with an accuracy rate of 96.67%.The difference over 1 year was 9.03% (with underestimation of 6.43% and overestimation of 2.60%), in which corresponding age data was of being less in training set or sesamoid nearby adductor pollicis or fusion of epiphysis appeared in test set.
Conclusion:An AI model based on deep learning of traditional ROIs′ features in hand DR images is initially achieved to automatically predict BA rapidly and effectively, yet it still needs further optimization.
