Performance comparison of different model configurations using the 3D Swin transformer backbone. Best BATE values are highlighted in bold
| Model | Predictive accuracy | Causal effect bias (BATE) () | |||
|---|---|---|---|---|---|
| MAE | MSE | ChemoRT versus RT | ChemoRT versus RT+EGFRI | RT versus RT+EGFRI | |
| Baseline (concatenation) | 1.75 0.08 | 4.21 0.25 | 1.82 0.15 | 0.41 0.09 | 1.63 0.18 |
| Baseline + Bi-AdaIN | 1.61 0.07 | 4.09 0.23 | 1.79 0.14 | 0.40 0.08 | 1.60 0.17 |
| Baseline + adversarial | 1.79 0.09 | 5.06 0.28 | 1.48 0.18 | 0.15 0.11 | 1.31 0.20 |
| Bi-AdaIN + adversarial | 1.64 0.08 | 4.26 0.26 | 1.41 0.15 | 0.15 0.09 | 1.24 0.16 |
| Baseline + MI | 1.80 0.09 | 5.19 0.29 | 0.35 0.12 | 0.09 0.07 | 0.35 0.14 |
| Bi-AdaIN + MI (proposed) | 1.62 0.07 | 4.13 0.24 | 0.140.05 | 0.060.04 | 0.110.06 |
| Model | Predictive accuracy | Causal effect bias ( | |||
|---|---|---|---|---|---|
| ChemoRT versus | ChemoRT versus RT+EGFRI | ||||
| Baseline (concatenation) | 1.75 | 4.21 | 1.82 | 0.41 | 1.63 |
| Baseline + Bi-AdaIN | 1.61 | 4.09 | 1.79 | 0.40 | 1.60 |
| Baseline + adversarial | 1.79 | 5.06 | 1.48 | 0.15 | 1.31 |
| Bi-AdaIN + adversarial | 1.64 | 4.26 | 1.41 | 0.15 | 1.24 |
| Baseline + | 1.80 | 5.19 | 0.35 | 0.09 | 0.35 |
| 1.62 | 4.13 | ||||
The BATE values involving the RT+EGFRI treatment group should be interpreted with caution due to the smaller sample size (n = 72) in this cohort compared to the ChemoRT (n = 1,413) and RT (n = 1,861) groups