Table 3 Performance of artificial intelligence algorithms in identifying implants for total hip arthroplasty
Author | AI Technique | DCNN | AUC | Accuracy | Sensitivity/Recall | Precision/PPV | Specificity | Processing Speed per Radiograph |
|---|---|---|---|---|---|---|---|---|
Borjali et al., 2020 [24] | DCNN | DenseNet-201 | NR | 100% | NR | NR | NR | NR |
Borjali et al., 2021 [25] | DCNN | DenseNet-201 | NR | 78% Human: 85% | NR | NR | NR | NR |
Gong et al., 2022 [27] | CNN Transfer Learning Framework Backward-Propagation Hyperparameter Tuning Data Augmentation | ResNet-50 | NR | Stem network: 91.5% Cup network: 83.7% Combined: 88.6% Joint network: 88.8% | Stem Network: 84.7% Cup Network: 75.4% Combined: 76.5% Joint Network: 82.1% | Stem Network: 91.5% Cup Network: 83.7% Combined: 88.6% Joint Network: 88.8% | NR | NR |
Kang et al., 2020 [29] | Image Augmentations Histogram Equalization Flipping Rotating | Keras API | 0.99 | NR | NR |  > 99% | NR | NR |
CNN Class Activation Heatmap | InceptionV3 | 0.999 | 99.60% | 94.3% | NR | 99.8% | NR | |
Karnuta et al., 2022 [21] | Image Preprocessing CNN Development | CNN | 0.999 | 99.6 | 94.3% | 93.6% | 99.8% | 0.02Â s |
Klemt et al., 2022 [30] | CNN Preprocessing Hyperparameter Optimization, Class Activation Heat Maps | InceptionV3 | NR | Primary THA: 98.2% Revision THA: 98.0% | Primary THA: 95.8% Revision THA: 94.9% | NR | Primary THA: 98.6% Revision THA: 98.0% | NR |
Murphy et al., 2022 [36] | Dropout and Batch Normalization Techniques | DenseNet-201 | NR | 91.7% | NR | NR | NR | 0.96 ± 0.02 s |
Patel et al., 2021 [31] | DCNN, Hyperparameter Optimization, Image Segmentation/Data Augmentation Ensembled Networks | EfficientNet & U-Net | NR | 98.9% Human: 76.1% | 98.90% | 99% | NR | 0.06 s vs Surgeon: 8.4 ± 6.1 min |
Median (IQR) | NA | NA | 0.999 (0.995–0.999) | 98.2% (91.7%–99.6%) | 94.6% (94.3%–95.7%) | 96.3% (93.1%–99.0%) | 99.2% (98.5%–99.8%) | 0.06 (0.04–0.51) |