Skip to main content

Table 3 Summary of the bone age testing methods based on deep learning

From: Fine-grained precise-bone age assessment by integrating prior knowledge and recursive feature pyramid network

Category of the method References Method Dataset Resolution of label MAE (month)
End-to-end Alexander et al. [28] Inception-V3 + DenseNet; fine-tuned ResNet-50; ice module; U-Net segmentation and CNN-based recognition RSNA dataset
Total data: 14,236
Training set: validation set: test set = 12,611:1425:200
The ratio of male to female is close to 1:1
1 month 4.27–4.5 months
Ren et al. [3] Regression CNN based on inception-V3 RSNA dataset
SCH data set (Shanghai Children’s Hospital): total data: 12,390
Training set: validation set: test set = 9912:1239:1239
1 month 5.2 months
Chen et al. [10] ResNet + spatial transformer + LBP + GLCM + SVM Total data: 12,536 from Shengjing Hospital, Training set: test set = 85:15 Not mentioned 5.5 months
Mutasa et al. [29] ResNet and inception Total data: 20,581 (10,289 from the network, 8909 from Columbia University Hospital, 1383 from the public data set)
Training set: validation set: test set = 11,007:1105:300
1 month 6.43 months
Liu et al. [30] NSCT-based multi-scale CNN (VGG16) Digital Hand Atlas Database System
Total data: 1391
12 months (1 year) 6.43 months
Iglovikov et al. [47] U-Net + VGG-style + linear combination RSNA dataset 1 month 7.52 months
Hu et al. [48] AlexNet DR images of Uygur people, total data: 472
Training set: test set = 7:3
1 month 8.40 months
Lee et al. [16] Caffenet architecture Total data: 600
Training set: test set = 2:1
Not mentioned 18.9 months
Synho et al. [18] CNN segmentation + Lenet BAA Training set: verification set: test set = 70:15:15 12 months (1 year) Accuracy of MAE < 1 year: 92.29%
Tajmir et al. [17] LeNet-5 Total data: 8325
Training set: test set = 8045:280
12 months (1 year) Accuracy of MAE < 1 year: 98.6%
Subregion-based Son et al. [32] faster R-CNN + VGG Total data: 3344
Training set: test set = 4:1
Age: 2–14 years old
Not mentioned 5.52 months
Reject rate: 1.6%
Bui et al. [20] faster R-CNN + TW3 Total data: 1375
Training set: test set = 4:1
12 months (1 year) 7.08 months
Spampinato et al. [9] BoNet (ad-hoc CNN, 6 layers) Digital Hand Atlas Database System
Total data: 1391
12 months (1 year) 9.48 months
Proposed method U-Net + RPN + cascade R-CNN + DenseNet with self-attention Total data: 2129
Annotated for fine-grained BAA: 873
Training set: test set = 4:1
Age: 2–18 years
12 months (1 year) 7.32 months
(7.08 months for age 2–18)
Reject rate: 0