Table 3. The P-values of the one-tailed t-test comparing the logarithm of χ2 values between the model meeting a given pre-selection criteria (applying both the 4D COV and the combined univariate regression P-value and FDR adjusted Q-value cutoffs) and its corresponding model without 4D COV pre-selection (applying the combined univariate regression P-value and FDR adjusted Q-value cutoffs alone).
A 3-feature radiomic model developed from the stable features with 4D COV≤5% and univariate regression P-value<0.01 and Q-value<0.05 achieved the most significantly improved performance over its counterpart without 4D COV pre-selection (i.e. developed from all features with univariate regression P-value<0.01 and Q-value<0.05 regardless of their 4D COV values), with a P-value of 2.16x10-27.
| Pre-selection criteria | Univariate P<0.005, Q<0.04 | Univariate P<0.007, Q<0.045 | Univariate P<0.01, Q<0.05 | Univariate P<0.03, Q<0.07 | Univariate P<0.05, Q<0.09 |
|---|---|---|---|---|---|
| 4D COV≤5% | 6.07 x10-8 | 1.28 x10-16 | 2.16 x10-27 | 1.66 x10-7 | 0.037 |
| 4D COV≤10% | 7.93 x10-3 | 9.84 x10-4 | 2.23 x10-4 | 1.94 x10-2 | 0.463 |
| 4D COV≤15% | 7.93 x10-3 | 9.84 x10-4 | 2.23 x10-4 | 1.974 x10-2 | 0.302 |
| 4D COV≤20% | 0.258 | 0.137 | 3.34 x10-2 | 0.154 | 0.396 |
| 4D COV≤25% | 0.500 | 0.500 | 0.500 | 0.087 | 0.016 |