Table 1.

Performance of NoduleX models.

Model	auc	S1 vs S45		spc	auc	S12 vs S45		spc	S0 vs S1-5		sens	spc
		acc	sens			acc	sens		auc	acc
CNN47	0.974	0.913	0.885	0.942	0.938	0.879	0.879	0.879	0.949	0.899	0.877	0.920
CNN47 + RF	0.993	0.952	0.942	0.962	0.943	0.894	0.864	0.924	0.984	0.946	0.948	0.943
CNN21	0.966	0.913	0.962	0.865	0.929	0.886	0.864	0.909	0.945	0.880	0.835	0.925
CNN21 + RF	0.989	0.962	0.962	0.962	0.971	0.932	0.879	0.985	0.975	0.925	0.906	0.943
LM	0.963	0.885	0.865	0.904	0.940	0.826	0.697	0.955	0.689	0.538	0.358	0.972

The performance of the two CNN models (CNN47: 47 × 47 × 5 and CNN21: 21 × 21 × 5) is shown with and without the addition of QIF features (CNN47 + RF, CNN21 + RF). Each model was tested on the validation set for three datasets: S1 vs S45, S12 vs S45, and S0 vs S1-4 (“non-nodule vs nodule”). Also shown is a simple logistic regression model based on the square root of the nodule’s greatest cross-sectional area (LM) for a baseline comparison. All models are measured on area under the ROC curve (auc), accuracy (acc), sensitivity (sens), and specificity (spc). The best performance for each metric is shown in bold.