Table 4.
Comparisons of classification-based accuracy between the SVM and CNN classifiers for six subregions of ILD with the same augmented data set. From each original 30 × 30 ROI patch, 10 ROI patches of 20 × 20 pixels were cropped from the five fixed positions (top-left, top-right, center, bottom-left, and bottom-right) and all the cropped patches were horizontally flipped. Both of the classifiers used the same data set augmented 10 times
| Study | Classification accuracy (%) | p value | |
|---|---|---|---|
| SVM | CNN | ||
| intra.G | 88.93 ± 1.19 | 95.86 ± 1.29 | < 0.001 |
| intra.S | 90.31 ± 1.00 | 95.21 ± 1.60 | < 0.001 |
| Integ | 86.81 ± 0.84 | 94.89 ± 1.60 | < 0.001 |
| inter.G2S | 75.90 ± 1.57 | 86.04 ± 2.22 | < 0.001 |
| inter.S2G | 73.11 ± 1.49 | 84.67 ± 2.50 | < 0.001 |