Table 1.
Ablation study on the ECG-IVUS dataset based on GoP recall.
| % | No training augmentation | Training augmentation | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Backbone | Bi-GRU | Bi-LSTM | Bi-GRU | Bi-LSTM | |||||||||
| TTEA | None | Tem. | Both | None | Tem. | Both | None | Tem. | Both | None | Tem. | Both | |
| w/o att. | Fold 1 | 69.08 | 72.38 | 72.58 | 68.75 | 69.02 | 70.93 | 66.51 | 66.58 | 65.59 | 72.58 | 73.37 | 74.75 |
| Fold 2 | 63.17 | 64.46 | 63.56 | 63.96 | 62.48 | 64.85 | 59.60 | 62.38 | 63.47 | 64.55 | 65.54 | 66.83 | |
| Fold 3 | 57.92 | 58.89 | 60.45 | 57.92 | 59.38 | 58.50 | 64.24 | 62.59 | 63.65 | 62.59 | 62.49 | 62.78 | |
| All | 64.17 | 66.23 | 66.51 | 64.26 | 64.37 | 65.61 | 63.89 | 64.23 | 64.43 | 67.41 | 68.00 | 69.04 | |
| with att. | Fold 1 | 71.00 | 72.64 | 71.85 | 70.73 | 71.32 | 70.67 | 72.58 | 73.83 | 73.70 | 70.80 | 69.94 | 72.18 |
| Fold 2 | 66.44 | 66.14 | 67.62 | 67.33 | 66.24 | 66.24 | 66.63 | 65.64 | 66.24 | 67.33 | 69.01 | 69.31 | |
| Fold 3 | 60.54 | 62.49 | 62.10 | 60.16 | 61.03 | 63.46 | 62.20 | 66.96 | 66.67 | 65.60 | 66.67 | 67.25 | |
| All | 66.68 | 67.86 | 67.83 | 66.70 | 66.90 | 67.32 | 67.89 | 69.52 | 69.54 | 68.31 | 68.73 | 69 . 94 | |
Test-time embedding augmentation (TTEA) is not applied (None), applied on the temporal difference encoded features (Tem.) or on both temporal difference encoded and spatiotemporal rotation encoded features (Both). The model performance with or without attention (Att.) are both evaluated. The bold value indicates the highest accurate rate in all experiments.