| | Publication reference | Region of interest | Modality (imaging sequence) | Data set | Network architecture | Performance |
| | Kompella et al. [37] | FC | Ultrasound | 256 images (training : validation: 85% : 15%) | Mask R-CNN | DSC: 0.80 (FC) |
| | Norman et al. [46] | FC, lateral TC, medial TC, PC, lateral menisci, medial menisci | MRI (T1-weighted, DESS) | OAI: 174 images (121 training, 37 validation, 16 testing) | U-Net | DSC (T1-weighted): 0.742 (FC, lateral TC, medial TC, PC), 0.767 (lateral menisci, medial menisci) | | DSC (DESS): 0.867 (FC, lateral TC, medial TC, PC), 0.833 (lateral menisci, medial menisci) |
| | Si et al. [47] | FC, TC, PC | MRI (sagT1-weighted, sagT2-weighted, corPDW FS, transversal PDW FS) | Tongren Hospital: 47 subjects (27 training, 20 testing) | U-Net | DSC± SD: 0.87 ± 0.01 (FC), 0.82 ± 0.01(TC), and 0.76 ± 0.04 (PC) | | Wirth et al. [31] | Medial FC, lateral FC, medial TC, lateral TC | MRI (corFLASH, sagDESS) | OAI: 92 subjects (50 training, 21 validation, 21 testing) | U-Net | DSC± SD (corFLASH): 0.92 ± 0.02 (medial TC), 0.88 ± 0.03 (medial FC), 0.92 ± 0.02 (lateral TC), 0.88 ± 0.02 (lateral FC) | | DSC± SD (sagDESS): 0.91 ± 0.02 (medial TC), 0.89 ± 0.03 (medial FC), 0.92 ± 0.02 (lateral TC), 0.90 ± 0.02 (lateral FC) |
| | Prasoon et al. [48] | TC | MRI (turbo 3D-T1-weighted) | (25 training, 114 testing) images | Three 2D CNN | DSC: 0.8249 (TC); SN: 81.92% (TC); SP: 99.97% (TC) | | Panfilov et al. [36] | FC, TC, PC, menisci | MRI (DESS) | OAI: 88 subjects | U-Net-mixup-unsupervised domain adaptation | DSC±SD: 0.907 ± 0.019 (FC), 0.897 ± 0.028 (TC), 0.871 ± 0.046 (PC), 0.863 ± 0.034 (menisci) | | Byraet al. [32] | Menisci | MRI (3D UTE cones) | University of California | 2D attention U-Net | DSC: 0.860 (menisci) | | San Diego Institutional Review Board: 61 subjects (36 training, 10 validation, 15 testing) |
| | Gajet al. [49] | FC, lateral TC, medial TC, PC, lateral menisci, medial menisci | MRI (3D-DESS) | OAI: 176 images (122 training, 36 validation, 18 testing) | U-Net-conditional generative adversarial networks | DSC± SD: 0.8972 ± 0.023 (FC), 0.9181 ± 0.013 (lateral TC), 0.8609 ± 0.038 (medial TC), 0.8417 ± 0.058 (PC), 0.8950 ± 0.023 (lateral menisci), 0.8738 ± 0.045 (medial menisci) | | Liu et al. [50] | FC, TC, FB, TB | MRI (T1-weighted SPGR) | SKI10: (60 training, 40 testing) images | SegNet + 3D simplex deformable modelling | ASD ± SD: 0.56 ± 0.12 mm (FB), 0.50 ± 0.14 mm (TB) | | VOE = 28.4 (FC), 33.1(TC) |
| | Zhou et al. [51] | FC, TC, PC, FB, TB, PB, menisci | MRI (3D-FSE) | 60 images | SegNet + conditional random field + 3D simplex deformable model | DSC ± SD: 0.97 ± 0.01 (FB), 0.962 ± 0.015 (TB), 0.898 ± 0.033 (PB), 0.806 ± 0.062 (FC), 0.801 ± 0.052 (TC), 0.807 ± 0.101 (PC), 0.831 ± 0.031 (menisci) |
|
|
Note. Region of interest: femoral cartilage (FC), tibial cartilage (TC), patellar cartilage (PC), femur bone (FB), tibia bone (TB), and patella bone (PB); modality (imaging sequence): magnetic resonance imaging (MRI); data set: Osteoarthritis Initiative (OAI); network architecture: convolutional neural network (CNN); performance: Dice similarity coefficient (DSC), specificity (SP), sensitivity (SN), average symmetric surface distance (ASD), and standard deviation (SD).
|
