Table 5 Quantitative performance comparison between MRI-based liver parenchyma and liver veins segmentation methods.

	Model	Liver parenchyma	Portal veins	Hepatic veins		Dataset
		Mean ± SD	Median (IQR)	Median (IQR)	Contrast enhanced	MRI characteristrics	Subjects	Name	Test size, n
Kavur et al.¹⁴	nnU-Net	0.954 ± 0.01	-	-	No	T1-DUAL in-phase, opposed-phase	healthy	CHAOS	20
Kart et al.¹³	nnU-Net	0.972 ± 0.02	-	-	No	T1 Dixon water	healthy	UKBB	200
Kart et al.¹³	nnU-Net	0.984 ± 0.01	-	-	No	T1 Dixon in-phase, opposed-phase, water, fat	healthy	GNC	200
Ivashchenko et al.²⁵	workflow	0.950 ± 0.01	-	-	Yes	multiphase T1 mDixon water	lesions	private	15
Ivashchenko et al.²⁶	DVNet	-	0.603 (0.08)	0.647 (0.05)	Yes	T1 mDixon	tumors	private	20
Ours	nnU-Net	0.936 ± 0.02	0.659 (0.11)	0.548 (0.16)	No	T1 Dixon in-phase, opposed-phase, water, fat	lesions	private	30

Summary of papers on MRI liver and vessels segmentation. Our study is the first to evaluate a 3D convolutional neural network (nnU-Net²³) for automated segmentation on non-contrast T1 vibe Dixon liver MRIs with lesions.
All scores are in Dice metric. Liver parenchyma results are shown as mean ± SD. The portal veins and hepatic veins segmentation results are compared using median and interquartile range (IQR). The MRI dataset used by each method is summarized.
MRI, magnetic resonance imaging; CHAOS, combined healthy abdominal organ segmentation challenge; UKBB, UK Biobank; GNC, German National Cohort.

Quick links

Search