Table 1.
Quantitative evaluation of the hybrid CNN-ViT networks compared to the corresponding pure CNN and transformer networks. Negative RVD values indicate a predicted volume smaller than the reference volume, whereas positive RVD values indicate a predicted volume larger than the reference volume
| Prediction on the test set | |||||||
|---|---|---|---|---|---|---|---|
| Organ | Model | DSC (%) | HD95 (mm) | RVD (%) | ASD (mm) | Sensitivity (%) | p-value(DSC) |
| Prostate | ViT | 83.91 ± 0.63 | 11.89 ± 1.65 | −16.40 ± 1.42 | 3.24 ± 1.71 | 81.04 ± 1.36 | < 0.001* compared to ResNet50-UNet-ViT and VGG16-UNet-ViT |
| ResNet50-UNet | 86.01 ± 1.84 | 3.11 ± 2.09 | −12.22 ± 2.91 | 1.38 ± 0.71 | 85.90 ± 1.02 | 0.014* compared to ResNet50-UNet-ViT | |
| ResNet50-UNet-ViT | 90.02 ± 1.00 | 2.85 ± 1.46 | −6.34 ± 1.78 | 0.96 ± 0.97 | 89.55 ± 1.45 | ||
| VGG16-UNet | 89.16 ± 1.03 | 2.44 ± 1.21 | + 0.71 ± 1.44 | 0.91 ± 0.81 | 87.99 ± 0.69 | <0.001* compared to VGG16-UNet-ViT | |
| VGG16-UNet-ViT | 91.75 ± 1.36 | 2.00 ± 1.11 | + 1.23 ± 1.02 | 0.53 ± 0.24 | 91.10 ± 1.00 | ||
| Bladder | ViT | 91.22 ± 0.94 | 1.41 ± 0.86 | + 2.37 ± 0.84 | 0.84 ± 0.69 | 91.32 ± 0.94 | < 0.001* compared to ResNet50-UNet-ViT and 0.001* compared to VGG16-UNet-ViT |
| ResNet50-UNet | 94.24 ± 1.02 | 1.51 ± 0.86 | + 1.00 ± 0.79 | 0.42 ± 0.84 | 93.99 ± 1.84 | 0.077 compared to ResNet50-UNet-ViT | |
| ResNet50-UNet-ViT | 94.98 ± 0.83 | 1.32 ± 0.80 | + 2.44 ± 1.37 | 1.04 ± 0.84 | 94.67 ± 1.03 | ||
| VGG16-UNet | 94.04 ± 1.23 | 1.89 ± 0.63 | + 2.42 ± 0.78 | 0.85 ± 0.81 | 94.24 ± 1.71 | 0.001* compared to VGG16-UNet-ViT | |
| VGG16-UNet-ViT | 95.32 ± 0.96 | 1.30 ± 0.99 | + 0.06 ± 1.01 | 0.64 ± 0.61 | 95.01 ± 0.87 | ||
| Rectum | ViT | 80.46 ± 1.09 | 8.63 ± 1.22 | −4.47 ± 0.96 | 3.49 ± 0.86 | 80.11 ± 1.23 | 0.016* compared to ResNet50-UNet-ViT and 0.001* compared to VGG16-UNet-ViT |
| ResNet50-UNet | 83.56 ± 1.45 | 4.23 ± 1.46 | + 1.10 ± 0.88 | 1.00 ± 0.57 | 83.20 ± 1.32 | 0.098 compared to ResNet50-UNet-ViT | |
| ResNet50-UNet-ViT | 83.86 ± 1.69 | 3.25 ± 1.11 | −2.03 ± 1.04 | 0.87 ± 1.01 | 83.07 ± 0.91 | ||
| VGG16-UNet | 85.01 ± 1.22 | 3.11 ± 0.89 | −5.17 ± 1.06 | 0.63 ± 1.22 | 84.45 ± 0.90 | 0.001* compared to VGG16-UNet-ViT | |
| VGG16-UNet-ViT | 87.00 ± 1.97 | 4.46 ± 0.94 | −1.49 ± 1.54 | 0.21 ± 0.88 | 86.52 ± 1.00 | ||
| RFH | ViT | 94.06 ± 0.63 | 1.72 ± 1.10 | + 1.30 ± 1.14 | 0.61 ± 0.82 | 93.29 ± 0.84 | 0.031* compared to ResNet50-UNet-ViT and 0.022* compared to VGG16-UNet-ViT |
| ResNet50-UNet | 95.14 ± 0.97 | 1.61 ± 0.94 | + 1.88 ± 0.97 | 0.27 ± 0.64 | 95.16 ± 1.05 | 0.011* compared to ResNet50-UNet-ViT | |
| ResNet50-UNet-ViT | 95.83 ± 0.95 | 1.22 ± 0.92 | −1.52 ± 1.42 | 0.81 ± 0.95 | 94.90 ± 1.12 | ||
| VGG16-UNet | 95.75 ± 0.91 | 1.36 ± 1.51 | + 1.37 ± 0.84 | 0.49 ± 0.60 | 95.09 ± 0.55 | 0.025* compared to VGG16-UNet-ViT | |
| VGG16-UNet-ViT | 96.30 ± 0.65 | 1.30 ± 1.01 | −1.15 ± 1.25 | 0.38 ± 0.68 | 96.39 ± 0.69 | ||
| LFH | ViT | 93.87 ± 1.45 | 1.89 ± 0.90 | + 1.51 ± 0.81 | 1.10 ± 0.80 | 93.39 ± 1.01 | 0.019* compared to ResNet50-UNet-ViT and 0.004* compared to VGG16-UNet-ViT |
| ResNet50-UNet | 95.02 ± 1.00 | 1.40 ± 1.31 | −4.01 ± 1.24 | 0.39 ± 0.56 | 94.26 ± 0.84 | 0.147 compared to ResNet50-UNet-ViT | |
| ResNet50-UNet-ViT | 94.80 ± 0.97 | 1.41 ± 0.86 | + 2.47 ± 0.91 | 0.51 ± 1.11 | 93.92 ± 1.24 | ||
| VGG16-UNet | 95.72 ± 0.59 | 1.60 ± 1.04 | + 0.69 ± 0.42 | 1.03 ± 1.26 | 95.08 ± 0.86 | 0.009* compared to VGG16-UNet-ViT | |
| VGG16-UNet-ViT | 96.34 ± 0.63 | 1.24 ± 1.44 | −0.72 ± 0.92 | 0.49 ± 1.39 | 96.22 ± 0.98 | ||
The best performance for each organ is highlighted in bold