Table 3.

Koos classification results obtained from automatic methods and human annotators using 5-fold cross-validation.

	MA-MAE	F1 score (%)	Accuracy (%)
ceT1
Baseline	0.23 ± 0.11	76.1 ± 6.3	76.6 ± 5.6
DenseNet	0.17 ± 0.04	81.5 ± 5.1	81.5 ± 5.1
Random forest	0.12 ± 0.05	87.6 ± 3.0	87.6 ± 3.0
hrT2
Baseline	0.22 ± 0.05	79.4 ± 2.1	79.5 ± 2.0
DenseNet	0.15 ± 0.06	83.8 ± 5.8	83.8 ± 5.8
Random forest	0.14 ± 0.06	85.2 ± 4.8	85.2 ± 4.8
ceT1 + hrT2
Baseline	0.22 ± 0.05	77.1 ± 3.7	77.2 ± 3.8
DenseNet	0.18 ± 0.05	82.1 ± 5.0	82.1 ± 5.0
Random forest	0.12 ± 0.06	87.2 ± 2.8	87.2 ± 2.8
Ensemble
DenseNet +	0.11 ±0.05	89.3 ±3.0	89.3 ±2.9
Random forest
Human annotators
Annotator 1	0.17 ± 0.07	85.4 ± 4.0	84.4 ± 4.7
Annotator 2	0.06 ± 0.02	92.9 ± 3.2	92.9 ± 3.1
Average human annotator	0.11 ± 0.08	89.1 ± 5.2	88.6 ± 5.8

The columns correspond to the weighted macro-averaged mean absolute error (MA-MAE), the weighted macro-averaged F1 score, and the accuracy score. Inputs are contrast-enhanced T1-weighted (ceT1) images, high-resolution T2-weighted (hrT2) images, or a combination of both (ceT1 + hrT2). The error ranges correspond to the standard deviation of the mean of values obtained from 5-fold cross-validation. Bold scores indicate the best automatic method.