Retrospective Study
Copyright ©The Author(s) 2022.
World J Orthop. Jun 18, 2022; 13(6): 603-614
Published online Jun 18, 2022. doi: 10.5312/wjo.v13.i6.603
Table 1 Data splits in training, testing and validation subsets according to Kellgren-Lawrence grades
Osteoarthritis Kellgren-Lawrence grade
Training
Testing
Validation
Samples, n
Proportion, %
Samples, n
Proportion, %
Samples, n
Proportion, %
025517.63717.67317.7
121314.73114.76114.8
216411.42411.44711.4
323716.43516.76816.4
457639.98339.516439.7
Total1445100210100413100
Table 2 Evaluation of parameters for knee osteoarthritis detection
Parameter

AccuracyDetermines the accuracy of the standalone model inaccuracy to detect the presence of KOA and its classification in the input image
PrecisionTrue positive/true positive + false positive
RecallTrue positive/true positive + false negative
LossDetermines the loss of the model
Table 3 Performance comparison of various transfer learning convolutional neural network models and eight expert human interpretations used for the development of deep learning algorithm for orthopedic radiographs
Model name
Accuracy
Precision
Recall
Loss
Outcome
ResNet5054.29%61.03%39.52%1.06Average
VGG-1656.68%67.56%35.02%1.10Average
InceptionV387.34%89.19%85.67%0.35Good
MobilnetV282.15%84.66%80.21%0.46Average
EfficientnetB756.61%70.09%38.27%0.98Average
DenseNet20192.87%93.69%92.53%0.20Best
Xception82.81%85.03%77.05%0.50Average
NasNetMobile80.90%83.98%77.30%0.50Average
Surgeon74.22%79.50%50.00%0.25Good