Table 1 Comparison of the the segmentation accuracy between different models in terms of the DSC, MSD and 95p HD (mean ± SD and median) for test set 1 annotated by observer 1.

Organ	Algorithm	Model	DSC		MSD (mm)		95p HD (mm)
Organ	Algorithm	Model	Mean ± SD	Median	Mean ± SD	Median	Mean ± SD	Median
Heart	nnU-Net	3d_fullres	0.95 ± 0.01	0.950	0.08 ± 0.02	0.088	0.30 ± 0.07	0.28
		3d_lowres	0.95 ± 0.01	0.949*	0.09 ± 0.02	0.090*	0.30 ± 0.07	0.28
		3d_cascade	0.95 ± 0.01	0.950	0.09 ± 0.02	0.089	0.31 ± 0.07	0.28
		2d	0.94 ± 0.01	0.945*	0.10 ± 0.01	0.096*	0.33 ± 0.06	0.31*
	AIMOS	UNet-768	0.94 ± 0.01	0.945*	0.10 ± 0.02	0.094*	0.33 ± 0.07	0.31*
Spinal Cord	nnU-Net	3d_fullres	0.91 ± 0.02	0.913	0.04 ± 0.01	0.034	0.28 ± 0.02	0.28
		3d_lowres	0.90 ± 0.02	0.898*	0.04 ± 0.01	0.041*	0.29 ± 0.03	0.28
		3d_cascade	0.91 ± 0.01	0.912	0.03 ± 0.01	0.034	0.28 ± 0.02	0.28
		2d	0.91 ± 0.02	0.907*	0.04 ± 0.01	0.036	0.28 ± 0.02	0.28
	AIMOS	UNet-768	0.90 ± 0.02	0.909*	0.04 ± 0.01	0.036*	0.28 ± 0.02	0.28
Right Lung	nnU-Net	3d_fullres	0.97 ± 0.01	0.970	0.04 ± 0.01	0.036	0.42 ± 0.00	0.42
		3d_lowres	0.97 ± 0.01	0.967*	0.04 ± 0.01	0.040*	0.42 ± 0.00	0.42
		3d_cascade	0.97 ± 0.01	0.970	0.04 ± 0.01	0.035	0.42 ± 0.00	0.42
		2d	0.97 ± 0.01 (0.97 ± 0.01)	0.966*	0.60 ± 0.84 (0.04 ± 0.01)	0.112*	5.29 ± 19.8 (0.42 ± 0.00)	0.42
	AIMOS	UNet-768	0.96 ± 0.01	0.963*	0.05 ± 0.01	0.044*	0.42 ± 0.00	0.42
Left Lung	nnU-Net	3d_fullres	0.97 ± 0.01	0.966	0.04 ± 0.01	0.035	0.53 ± 0.13	0.56
		3d_lowres	0.96 ± 0.01	0.962*	0.04 ± 0.01	0.042*	0.56 ± 0.00	0.56
		3d_cascade	0.97 ± 0.01	0.966	0.04 ± 0.01	0.035	0.53 ± 0.13	0.56
		2d	0.96 ± 0.01	0.965*	0.04 ± 0.02	0.037*	0.54 ± 0.09	0.56
	AIMOS	UNet-768	0.95 ± 0.01	0.956*	0.05 ± 0.01	0.045*	0.56 ± 0.00	0.56

The asterisk (*) indicates a significant difference with the nnU-Net 3d full resolution model according to the Wilcoxon signed rank test with a significance level of \(\alpha =0.05\). Values in parentheses correspond to re-calculated metrics after connected component analysis.

Quick links

Search