Table 8 Comparison of macro precision, recall and f1-scores for sequence classification using transformer and GRU, utilizing a trained ConvNeXt Tiny model for sequence embedding.
From: GastroHUN an Endoscopy Dataset of Complete Systematic Screening Protocol for the Stomach
Strategy | Training label | Transformer: macro | GRU: macro | ||||
|---|---|---|---|---|---|---|---|
precision | recall | f1-score | precision | recall | f1-score | ||
Consensus | All | 85.96 ± 0.47 | 86.34 ± 0.49 | 85.14 ± 0.48 | 85.49 ± 0.44 | 85.92 ± 0.44 | 84.86 ± 0.44 |
Triple | 81.46 ± 0.44 | 81.58 ± 0.45 | 80.51 ± 0.45 | 83.58 ± 0.44 | 83.17 ± 0.44 | 82.45 ± 0.43 | |
FG | 85.31 ± 0.36 | 84.14 ± 0.39 | 83.33 ± 0.40 | 85.59 ± 0.40 | 84.40 ± 0.41 | 83.66 ± 0.41 | |
G | 81.95 ± 0.45 | 81.34 ± 0.46 | 80.46 ± 0.45 | 86.74 ± 0.38 | 86.09 ± 0.39 | 85.47 ± 0.39 | |
FG1 - G1 | 86.21 ± 0.40 | 85.53 ± 0.45 | 84.81 ± 0.44 | 84.07 ± 0.44 | 83.27 ± 0.49 | 82.85 ± 0.47 | |
FG1 - G2 | 86.98 ± 0.42 | 87.01 ± 0.41 | 86.30 ± 0.42 | 86.15 ± 0.41 | 85.63 ± 0.39 | 85.01 ± 0.41 | |
FG2 - G1 | 83.83 ± 0.49 | 82.67 ± 0.49 | 82.03 ± 0.48 | 81.84 ± 0.50 | 81.52 ± 0.56 | 80.53 ± 0.51 | |
FG2 - G2 | 82.62 ± 0.42 | 83.77 ± 0.44 | 82.00 ± 0.44 | 78.38 ± 0.46 | 79.54 ± 0.45 | 77.53 ± 0.46 | |
Annotator | FG1 | 80.99 ± 0.46 | 80.43 ± 0.49 | 79.52 ± 0.48 | 79.04 ± 0.53 | 78.40 ± 0.57 | 77.32 ± 0.56 |
FG2 | 79.10 ± 0.45 | 79.35 ± 0.51 | 77.47 ± 0.44 | 76.79 ± 0.51 | 76.38 ± 0.58 | 74.37 ± 0.55 | |
G1 | 81.54 ± 0.44 | 80.68 ± 0.41 | 80.12 ± 0.39 | 82.03 ± 0.39 | 81.28 ± 0.44 | 80.59 ± 0.42 | |
G2 | 80.57 ± 0.52 | 80.27 ± 0.54 | 79.38 ± 0.51 | 78.67 ± 0.53 | 78.83 ± 0.57 | 77.53 ± 0.55 | |
