Table 9.
Results of Transfer Learning. Comparison of segmentation performance of the proposed cGAN and U-Net without and with transfer learning and testing on the SKI10 and OAI ZIB testing dataset. Highest network scores achieved for each tissue are in bold.
| Transfer Learning Results | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| SKI10 Testing | |||||||||||
| Network | Training | F Bone | T Bone | F Cartilage | T Cartilage | ||||||
| DSC | ASD | DSC | ASD | DSC | VOE | DSC | VOE | ||||
| cGAN | AMROA | 0.929 ± 0.040 | 3.726 ± 1.758 | 0.893 ± 0.069 | 3.368 ± 1.935 | 0.488 ± 0.093 | 67.19 ± 8.36 | 0.465 ± 0.114 | 69.01 ± 10.00 | ||
| SKI10 | 0.974 ± 0.013 | 1.445 ± 1.918 | 0.979 ± 0.007 | 0.527 ± 0.403 | 0.736 ± 0.058 | 41.49 ± 6.99 | 0.684 ± 0.070 | 47.58 ± 7.98 | |||
| SKI10 → AMROA | 0.938 ± 0.039 | 3.229 ± 1.776 | 0.929 ± 0.041 | 2.696 ± 2.326 | 0.544 ± 0.077 | 62.23 ± 7.45 | 0.480 ± 0.100 | 67.86 ± 8.89 | |||
| AMROA → SKI10 | 0.974 ± 0.012 | 1.280 ± 1.484 | 0.977 ± 0.010 | 0.802 ± 1.139 | 0.738 ± 0.059 | 41.19 ± 7.08 | 0.675 ± 0.071 | 48.65 ± 7.94 | |||
| U-Net | AMROA | 0.925 ± 0.038 | 1.856 ± 0.997 | 0.907 ± 0.055 | 1.868 ± 1.336 | 0.545 ± 0.082 | 62.16 ± 7.62 | 0.462 ± 0.112 | 69.26 ± 9.86 | ||
| SKI10 | 0.973 ± 0.015 | 0.756 ± 0.995 | 0.978 ± 0.008 | 0.254 ± 0.340 | 0.728 ± 0.058 | 42.42 ± 6.88 | 0.674 ± 0.066 | 48.85 ± 7.55 | |||
| SKI10 → AMROA | 0.943 ± 0.032 | 1.071 ± 0.682 | 0.936 ± 0.038 | 1.436 ± 1.083 | 0.576 ± 0.078 | 59.18 ± 7.86 | 0.456 ± 0.115 | 69.76 ± 9.93 | |||
| AMROA → SKI10 | 0.975 ± 0.013 | 0.440 ± 0.492 | 0.979 ± 0.007 | 0.258 ± 0.288 | 0.731 ± 0.056 | 42.08 ± 6.74 | 0.670 ± 0.070 | 49.19 ± 7.84 | |||
| OAI ZIB Testing | |||||||||||
| cGAN | AMROA | 0.939 ± 0.016 | 4.153 ± 1.962 | 0.914 ± 0.080 | 4.681 ± 3.197 | 0.611 ± 0.068 | 55.66 ± 7.10 | 0.601 ± 0.089 | 56.44 ± 9.14 | ||
| OAI ZIB | 0.985 ± 0.002 | 0.328 ± 0.123 | 0.985 ± 0.003 | 0.293 ± 0.072 | 0.895 ± 0.023 | 18.92 ± 3.64 | 0.839 ± 0.040 | 27.55 ± 5.90 | |||
| OAI ZIB → AMROA | 0.961 ± 0.009 | 1.786 ± 1.202 | 0.961 ± 0.018 | 4.426 ± 2.902 | 0.641 ± 0.071 | 52.41 ± 7.87 | 0.738 ± 0.055 | 41.23 ± 6.70 | |||
| AMROA → OAI ZIB | 0.985 ± 0.002 | 0.403 ± 0.268 | 0.985 ± 0.003 | 0.293 ± 0.068 | 0.897 ± 0.022 | 18.68 ± 3.57 | 0.837 ± 0.042 | 27.82 ± 6.19 | |||
| U-Net | AMROA | 0.934 ± 0.015 | 5.424 ± 2.799 | 0.915 ± 0.094 | 6.282 ± 3.647 | 0.643 ± 0.065 | 52.26 ± 7.03 | 0.626 ± 0.063 | 54.12 ± 6.74 | ||
| OAI ZIB | 0.985 ± 0.002 | 0.388 ± 0.169 | 0.984 ± 0.003 | 0.304 ± 0.079 | 0.896 ± 0.020 | 18.83 ± 3.19 | 0.837 ± 0.038 | 27.80 ± 5.57 | |||
| OAI ZIB → AMROA | 0.966 ± 0.006 | 1.244 ± 0.791 | 0.961 ± 0.017 | 1.880 ± 1.133 | 0.734 ± 0.046 | 41.83 ± 5.82 | 0.741 ± 0.058 | 40.83 ± 6.97 | |||
| AMROA → OAI ZIB | 0.985 ± 0.002 | 0.390 ± 0.361 | 0.985 ± 0.003 | 0.327 ± 0.127 | 0.893 ± 0.023 | 19.24 ± 3.64 | 0.838 ± 0.037 | 27.75 ± 5.50 | |||
SKI10/OAI ZIB → AMROA: Pretraining the network for 20 epochs on the SKI10/OAI ZIB dataset followed by network fine-tuning for 80 epochs on the AMROA dataset.
AMROA → SKI10/OAI ZIB: Pretraining the network for 20 epochs on the AMROA dataset followed by network fine-tuning for 80 epochs on the SKI10/OAI ZIB dataset.
Results are presented as mean ± standard deviation.
Abbreviations: FB – femoral bone, TB – tibial bone, FC – femoral cartilage, TC – tibial cartilage, DSC - Sørensen–Dice similarity coefficient, ASD – average surface distance, VOE – volumetric overlap error.