Table 3.
Model efficiency and edge deployment performance.
| Methods/configuration | Params (M) | FLOPs (G) | GPU Latency (ms) | CPU Latency (ms) | FPS | Memory (MB) | Power (W) |
|---|---|---|---|---|---|---|---|
| Model complexity comparison | |||||||
| U-Net | 31.0 | 54.3 | 18.7 | 142.5 | 53 | – | – |
| U-Net + + | 36.2 | 61.8 | 21.3 | 168.9 | 47 | – | – |
| Attention U-Net | 33.7 | 57.6 | 19.8 | 155.2 | 50 | – | – |
| TransUNet | 105.3 | 98.4 | 43.6 | 387.2 | 23 | – | – |
| Swin-UNet | 89.4 | 52.7 | 37.8 | 312.5 | 26 | – | – |
| UNeXt | 1.47 | 3.8 | 9.6 | 96.8 | 104 | – | – |
| UNETR + + | 89.5 | 67.3 | 41.2 | 358.7 | 24 | – | – |
| MedNeXt | 18.6 | 31.4 | 25.6 | 198.4 | 39 | – | – |
| USEANet | 3.64 | 0.79 | 14.2 | 118.3 | 70 | – | – |
| Wave-GMS | 2.60 | 4.7 | 13.5 | 108.6 | 74 | – | – |
| CEA-Net (Ours) | 0.83 | 2.1 | 11.8 | 89.4 | 85 | – | – |
| Edge device deployment (Jetson Nano) | |||||||
| CEA-Net (Batch = 1) | 0.83 | 2.1 | – | 52.3 | 19 | 512 | 5.4 |
| CEA-Net (Batch = 2) | 0.83 | 2.1 | – | 98.7 | 20 | 697 | 6.2 |
| CEA-Net (Batch = 4) | 0.83 | 2.1 | – | 191.5 | 21 | 1138 | 7.1 |
| CEA-Net (Batch = 8) | 0.83 | 2.1 | – | 367.8 | 22 | 2015 | 8.3 |
GPU: NVIDIA RTX 3090; CPU: Intel i9-10900 K; Edge Device: NVIDIA Jetson Nano Developer Kit. FPS denotes frames (images) processed per second, computed as the batch size divided by the corresponding per-batch latency. “–” indicates not applicable.