Table 2 Performance of deep learning segmentation models
From: Automatic segmentation of fat metaplasia on sacroiliac joint MRI using deep learning
Models | Internal cross-validation | External test set | ||||
|---|---|---|---|---|---|---|
DSC (%) | Precision (%) | Recall (%) | DSC (%) | Precision (%) | Recall (%) | |
2.5D-AttentionUNet (ours) | 81.86 ± 1.55 | 80.49 (80.16–80.82) | 85.5 (85.34–85.66) | 85.44 ± 6.09 | 85.83 (82.62–89.04) | 86.43 (81.10–91.76) |
2D-UNet | 76.97 ± 2.11** | 78.5 (78.25–78.75) | 78.24 (77.84–78.64) | 66.53 ± 19.37** | 69.14 (62.20–74.04) | 68.09 (59.71–76.50) |
3D-UNet | 80.05 ± 1.57 | 80.18 (79.71–81.55) | 82.4 (82.00–82.80) | 67.40 ± 20.84** | 85.12 (80.31–89.37) | 62.87 (52.46–71.26) |
ResUNet | 76.42 ± 1.92 ** | 76.49 (76.15–76.83) | 79.56 (79.31–79.81) | 66.41 ± 17.60** | 75.74 (69.66–79.58) | 65.66 (57.32–72.96) |
UNETR | 73.94 ± 2.68 ** | 79.64 (79.12–80.16) | 72.93 (72.51–73.35) | 57.81 ± 21.58** | 68.83 (61.45–77.27) | 55.26 (46.22–67.42) |
Attention U-Net | 79.73 ± 1.65 * | 79.95 (79.59–80.31) | 82.47 (82.12–82.82) | 65.54 ± 24.25* | 72.29 (63.05–76.22) | 67.95 (56.84–76.66) |