Table 3.

Quantitative comparison of all the methods discussed in Section 2. Trade-off between qualities of COVID-19 identification and ranking by severity is observed for segmentation-based methods. The proposed Multitask-Spatial-1 model yields the best identification results. Results are given as $mean\pm std$.

	ROC AUC (COVID-19 vs. $·$)				Spearman’s $\rho$	Dice Score
	vs. All others	vs. Normal	vs. Bac. Pneum.	vs. Nodules
Thresholding	$.51\pm 0.00$	$.68\pm 0.00$	$.46\pm 0.00$	$.45\pm 0.00$	$.92\pm 0.00$	$.42\pm 0.00$
3D U-Net	$.76\pm 0.02$	$.89\pm 0.02$	$.59\pm 0.01$	$.79\pm 0.03$	$.\mathbf{97}\pm 0.01$	$.\mathbf{65}\pm 0.00$
2D U-Net	$.78\pm 0.01$	$.93\pm 0.01$	$.62\pm 0.01$	$.79\pm 0.00$	$.\mathbf{97}\pm 0.00$	$.63\pm 0.00$
2D U-Net+	$.86\pm 0.01$	$.\mathbf{98}\pm 0.01$	$.68\pm 0.02$	$.91\pm 0.01$	$.80\pm 0.03$	$.59\pm 0.01$
ResNet-50	$.62\pm 0.19$	$.67\pm 0.21$	$.55\pm 0.13$	$.65\pm 0.22$	N/A	N/A
Multitask-Latent	$.79\pm 0.06$	$.84\pm 0.05$	$.73\pm 0.06$	$.80\pm 0.07$	$.\mathbf{97}\pm 0.00$	$.61\pm 0.02$
Multitask-Spatial-4	$.89\pm 0.03$	$.94\pm 0.03$	$.83\pm 0.05$	$.91\pm 0.03$	$.\mathbf{98}\pm 0.00$	$.61\pm 0.02$
Multitask-Spatial-1	$.\mathbf{93}\pm 0.01$	$.\mathbf{97}\pm 0.01$	$.\mathbf{87}\pm 0.01$	$.\mathbf{93}\pm 0.00$	$.\mathbf{97}\pm 0.01$	$.61\pm 0.02$