Study (Publication year) | AI name | CNN architecture type | ROI or Important region labeling | Data input proportion in training/validation/test |
---|---|---|---|---|
Adams [11] | DCNN | AlexNet or GoogLeNet architectural model | X | Training-57% Validation-29% Test-14% |
Urakawa [12] | CNN | VGG16 architectural model | X | Training-80% Validation-10% Test-10% |
Cheng [13] | DCNN | DenseNet-121 architectural model | Image labeling and preprocessing = Each image was reviewed by a trauma surgeon for the preciseness of the label and quality of the images | Training-60% Validation-20% |
Krogue [14] | Deep learning model | DenseNet containing 169 layers architectural model | Object detection algorithm to place the bounding boxes automatically = Single-shot detector with the Resnet-50 feature pyramid network architecture | Training-60% Validation-25% Test-15% |
Yu [15] | CNN | Inception-V3 architectural model | RoI identifying = Each ROI was either approved or revised by the local expert | Training-60% Validation-20% Test-20% |
Mutasa [16] | CNN | Novel 2D neural network utilizing a customized residual network based architecture | Highlight important regions = Gradient-weighted class activation mapping (Grad-CAM) | Training & validation-90% Test-10% |
Beyaz [17] | CNN | CNN containing GA blocks architectural model | Highlight important regions = Regions containing both fractured and non-fractured femoral necks were cropped from the X-ray images manually | X |
Mawatari [18] | DCNN | GoogLeNet architectural model | RoI identifying = All radiographs were manually checked and annotated retrospectively by the three radiologists referring to CT and MRI for RoI selection | Training-85% Test-15% |
Yamada [19] | CNN | Xception architectural model | Highlight important regions = Orthopedic surgeon (3 years of experience) performed the image preprocessing using Paint 3D (Microsoft Corp, Redmond, WA, USA) by cropping the minimum region containing the femoral head and greater and lesser trochanters | Training-95% Validation-5% |
Cheng [20] | DCNN | DenseNet-121 architectural model | Highlight important regions = Grad-CAM | Training-60% Validation-20% Test-20% |
Yoon [8] | Deep faster R-CNN | Math-Works (VGG-16 architecture) architectural model | Highlight important regions = Grad-CAM | Training-80% Test-20% |
Sato [1] | CNN | EfficientNet-B4 architectural model | Highlight important regions = Grad-CAM | Training-80% Validation-10% Test-10% |
Bae [21] | CNN | Modified spatial attention module (CBAM + +) and ResNet18 architectural model | Highlight important regions = Grad-CAM | Training-80% Validation-10% Test-10% |
Murphy [7] | CNN1 and CNN2 | GoogLeNet architectural model | RoI identifying = MATLAB Training Image Labeller Application (tool) | Training-60% Validation-20% Test-20% |