Table 2.
Impact of using an ensemble of CNNs on quantification for the pediatric lymphoma population. Shown are results for SUVmax, MTV, and TLG as summarized by Pearson’s correlation coefficient, R, and the relative percent difference when compared to physician measurements
SUVmax | MTV | TLG | ||||
---|---|---|---|---|---|---|
Method | R | Cases incorrect (of 100) |
R | RPD Median [IQR] |
R | RPD Median [IQR] |
1st CNN initialization | 0.73 | 24 | 0.89 |
− 3.5 [− 11.4, 4.2] |
0.95 |
0.4 [− 5.2, 7.7] |
2nd CNN initialization | 0.74 | 25 | 0.87 |
− 4 [− 11.7, 3.9] |
0.94 |
− 0.1 [− 5.1, 7] |
3rd CNN initialization | 0.59 | 24 | 0.89 |
− 3.8 [− 12.4, 5.3] |
0.95 |
0 [− 6.7, 7.6] |
Ensemble CNN | 0.86 | 19 | 0.88 |
− 4.2 [− 10, 5.7] |
0.94 |
− 0.4 [− 5.2, 7] |