Table 6.
Summary of 2D CNN classification approaches on grading of osteoarthritis severity.
| Publication reference | Target tasks | Modality (imaging sequence) | Data set | Network architecture | Performance |
|---|---|---|---|---|---|
| Tiulpin et al. [16] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | MOST: 18376 images (training), | Deep Siamese convolutional neural network | AUC: 0.93 |
| OAI: 2957 images (validation), 5960 images (testing) | |||||
|
| |||||
| Nguyen et al. [57] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | OAI: 39,902 images (training) | Deep Siamese convolutional neural network with pi-model approach | Cohen's Kappa coefficient (KC): 0.790 |
| MOST: 3,445 images (testing) | Balanced accuracy (BA): 0.527 | ||||
|
| |||||
| Nguyen et al. [23] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | OAI: 39902 images (training/validating) | Semixup (Siamese network + novel deep semisupervised learning) | Balanced accuracy ± SD: 71 ± 0.8% |
| MOST: 3445 images (testing) | |||||
|
| |||||
| Liu et al. [58] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | 2770 images | Faster R-CNN (region proposal network + Fast R–CNN) + focal loss | Accuracy: 82.5%; SN: 78.2%; SP: 94.8% |
| Antony et al. [30] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | OAI: 8892 images | VGG16, VGG-M-128, and BVLC | Mean squared error: 0.504 (CNN-Reg) |
| CaffeNet | |||||
|
| |||||
| Norman et al. [5] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | OAI: 39,593 images (25,873 training, 7779 validation, 5941 testing) | DenseNet | SN: 83.7 (no OA), 70.2 (mild OA), 68.9 (moderate OA), 86.0 (severe OA) % |
| SP: 86.1 (no OA), 83.8 (mild OA), 97.1 (moderate OA), 99.1 (severe OA) % | |||||
|
| |||||
| Zhang et al. [59] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | OAI: (38232 training, 10986 testing, 5422 validation) images | ResNet with convolutional block attention module (CBAM) | Accuracy: 74.81%; mean squared error: 0.36; quadratic Kappa score: 0.88 |
| Leung et al. [60] | Predict knee OA severity based on KL-grade and predict total knee replacement | X-ray (plain radiography) | OAI: 728 subjects | ResNet-34 (ResNet with 34 layers) | AUC: 0.87 |
| Tiulpin and Saarakkala [26] | Predict knee OA severity | X-ray (plain radiography) | OAI: 19704 images (training); MOST: 11743 (testing) | SE-ResNet-50 + SE-ResNet-50-32 × 4d (SE-ResNet-50 with ResNeXt blocks) | AUC: 0.98 |
| Kim et al. [17] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | 4366 images (3464 training, 386 validation, 516 testing) | Six SE-ResNet | AUC: 0.97 (KL 0), 0.85 (KL1), 0.75 (KL2), 0.86 (KL3), 0.95 (KL4) |
| Chen et al. [34] | Predict knee OA severity based on KL-grade | X-ray (plain radiography) | OAI: 4130 images (training : validation : testing; 7 : 1 : 2) | VGG-19 + proposed ordinal loss | Accuracy: 70.4%; mean absolute error (MAE): 0.358 |
| Pedoia et al. [61] | Predict presence of OA | MRI (T2 mapping acquisition) | OAI: 4384 subjects | DenseNet | AUC: 83.44%; SN: 76.99%; SP: 77.94% |
Note. Modality (imaging sequence): magnetic resonance imaging (MRI); data set: Osteoarthritis Initiative (OAI); network architecture: convolutional neural network (CNN) and squeeze-and-excitation (SE); performance: specificity (SP), sensitivity (SN), area under receiver operating characteristics curve (AUC), and standard deviation (SD).