Table 5.
Dynamic inference mechanism validation and performance under different cloud-edge modes.
| Mode/sample type | Dice (%) | Latency (ms) | Proportion (%) | Routing accuracy (%) |
|---|---|---|---|---|
| Deployment mode comparison | ||||
| Cloud-only mode | 88.3 ± 0.5 | 11.8 | – | – |
| Edge-only mode | 82.3 ± 0.8 | 6.5 | – | – |
| Cloud-edge collaborative mode | 87.1 ± 0.6 | 8.3 | – | 91.2 ± 1.0 |
| Dynamic inference sample analysis | ||||
| Simple samples (edge processing) | 88.9 ± 0.5 | 6.5 | 63.8 | – |
| Complex samples (cloud processing) | 84.2 ± 0.8 | 11.8 | 36.2 | – |
| Weighted average performance | 87.1 ± 0.6 | 8.3 | 100.0 | – |
All tests conducted on BUSI dataset. Cloud mode uses full 4-stage decoder, edge mode uses 3-stage early exit. Routing threshold θ = 0.7. Dice scores and routing accuracy are reported as mean ± standard deviation over five-fold cross-validation.