Table 3.
Experimental results on cryo-ET subtomogram segmentation using simulated and real datasets.
| Simulated dataset | ||
|---|---|---|
| Method | mIoU | Dice |
| U-CLSTM [SHREC21] | 76.8±0.4 | 87.2±0.1 |
| DeepFinder [SHREC21] | 81.0±0.1 | 89.5±0.1 |
| VoxResNet [Neurolmage18] | 84.3±0.4 | 91.4±0.2 |
| Swin UNETR [CVPR22] | 78.8±0.9 | 88.1±0.6 |
| MedNeXt [miccai23] | 82.7±0.5 | 90.8±0.3 |
| SwiFT [NeurIPS23] | 80.1±0.1 | 88.9±0.7 |
| MVGFormer(MLP) | 83.9±0.4 | 91.2±0.2 |
| MVGFormer(MF) | 84.9 ±0.3 | 91.9 ±0.2 |
| MVGFormer(P3DA) | 87.1 ±0.2 | 93.2 ±0.2 |
| PolyGA dataset | ||
| Method | mIoU | Dice |
| VoxResNet (w/o pre) | 46.0±1.5 | 61.6±0.9 |
| 3D-UNet [JBCB21] | 49.4±0.2 | 66.1±0.1 |
| DeepFinder [SHREC21] | 53.4±0.3 | 69.6±0.4 |
| MedNeXt [miccai23] | 55.2±0.5 | 71.1±0.2 |
| VoxResNet [Neurolmage18] | 59.8±0.6 | 74.8±0.4 |
| SwiFT [NeurIPS23] | 57.0±0.2 | 72.6±0.3 |
| MVGFormer(MLP) | 61.5±0.1 | 76.2±0.3 |
| MVGFormer(MF) | 62.3 ±0.1 | 76.8 ±0.4 |
| MVGFormer(P3DA) | 63.7 ±0.2 | 78.5 ±0.3 |
| Erwinia dataset | ||
| Method | mIoU | Dice |
| VoxResNet (w/o pre) | 72.4±0.7 | 80.3±1.1 |
| 3D-UNet [JBCB21] | 76.7±0.2 | 86.8±0.3 |
| DeepFinder [SHREC21] | 88.0±0.3 | 93.6±0.2 |
| MedNeXt [miccai23] | 89.4±1.1 | 94.4±0.1 |
| VoxResNet [Neurolmage18] | 95.4±0.1 | 97.6±0.1 |
| SwiFT [NeurIPS23] | 91.7±0.3 | 95.7±0.1 |
| MVGFormer(MF) | 96.8 ±0.2 | 98.4 ±0.2 |
| MVGFormer(P3DA) | 97.9 ±0.4 | 98.9 ±0.1 |
We use the bold and the underlined to represent the best results and the runner-ups, respectively.