Table 5 Results of Ablation Study of three variants (ResNet18, ResNet34, and ResNet50) of ResNet as our backbone network.
From: Emphysema subtyping on thoracic computed tomography scans using deep neural networks
Backbone | Subtype | Accuracy (%) | F-measure | Linear weighted kappa (95% CI) | #Param (M) | GMACs |
|---|---|---|---|---|---|---|
(a) Classification network in predicting paraseptal emphysema severity | ||||||
 ResNet18 | Centrilobular | 48.55 | 47.58 | 62.50 (61.83–63.63) | 34.48 | 1148.05 |
Paraseptal | 54.81 | 51.68 | 33.63 (31.77–35.49) | |||
 ResNet34 | Centrilobular | 52.33 | 51.00 | 64.29 (63.16–65.42) | 64.79 | 1661.95 |
Paraseptal | 59.12 | 57.12 | 42.03 (40.21–43.85) | |||
 ResNet50 | Centrilobular | 51.31 | 51.00 | 63.00 (61.84–64.16) | 47.86 | 1702.69 |
Paraseptal | 54.79 | 52.22 | 34.62 (32.76–36.48) | |||
(b) Regression network in predicting emphysema severity | ||||||
 ResNet18 | Centrilobular | 49.91 | 47.08 | 62.53 (61.39–63.67) | 34.48 | 1147.81 |
Paraseptal | 55.87 | 53.92 | 39.16 (37.40–40.92) | |||
 ResNet34 | Centrilobular | 51.32 | 49.61 | 64.24 (63.14–65.35) | 64.79 | 1661.72 |
Paraseptal | 64.62 | 60.74 | 52.06 (50.40–53.73) | |||
 ResNet50 | Centrilobular | 51.03 | 46.70 | 62.15 (60.98–63.31) | 47.86 | 1702.45 |
Paraseptal | 62.21 | 56.22 | 47.21 (45.45–48.97) | |||
