Emerging artificial intelligence applications in liver magnetic resonance imaging

Table 1 Applications of artificial intelligence segmentation methods to liver magnetic resonance imaging

Ref.	Task	Method	MR image	DICE
Mole et al[37], 2020	Segment liver from T1 mapping technique to aid surgical planning	3D U-Net	T1 map	0.97
Winther et al[27], 2020	Segment liver from Gd-EOB-DTPA-enhanced MRI for volume calculations	3D U-Net	Gadoxetic acid-enhanced MRI	0.96 ± 1.9
Liu et al[30], 2020	Segment liver for automated liver iron quantification	2D U-Net	T2* map	0.86 ± 0.01
Wang et al[43], 2019	Segment Liver across multiple imaging modalities and techniques	2D U-Net	T1- and T2*- weighted	T1-w: 0.95 ± 0.03
				T2-w: 0.92 ± 0.05
Cunha et al[46], 2020	Segment liver to classify if adequate contrast uptake has occurred in contrast enhanced scans	2D U-Net	Pre- and post-contrast T1- weighted, and T2- weighted	Not reported
Chen et al[31], 2020	Segment multiple organs in abdominal scans, to aid radiotherapy planning	2D Dense U-Net	T1-weighted	Liver: 0.96 ± 0.009
Bousabarah et al[36], 2020	Segment and delineate HCCs	2D U-Net	Gadoxetic acid-enhanced MRI	Liver: 0.91 ± 0.01
				Tumour: 0.68 ± 0.03
Ivashchenko et al[41], 2019	Segment liver, vasculature and biliary tree	4D k-mean clustering	Gadoxetic acid-enhanced MRI	Liver: 0.95 ± 0.01
Irving et al[44], 2017	Segment liver with vessel exclusion to assist in liver assessment	2D U-Net	T1 map	0.95
Yang et al[45], 2019	Segment liver across multiple domains via domain transfer	Cycle GAN and 2D U-Net	Gadoxetic acid-enhanced MRI	0.891 ± 0.040
Christ et al[39], 2017	Segment liver and tumours within, in CT and MRI	Two sequential 2D U-Nets	Diffusion-weighted	Liver: 0.87
				Tumour: 0.697
Fu et al[35], 2018	Segment multiple organs in abdominal scans, to aid radiotherapy planning	Three Dense CNNs	T2/T1-weighted	Liver: 0.953 ± 0.007
Valindria et al[33], 2018	Segment multiple organs in multi-modal (MR,CT) scans	ResNet Encoder Decoder	T2-weighted	Liver: 0.914
Masoumi et al[42], 2012	Segment the liver	Watershed (non-AI) + ANN	Abdominal MRI	0.94 (IoU not DICE)
Jansen et al[40], 2019	Segment liver and metastases	CNN	DCE-MR and diffusion-weighted	Liver: 0.95

MRI: Magnetic resonance imaging; CT: Computed tomography; DCE-MR: Dynamic contrast enhanced magnetic resonance; HCC: Hepatocellular carcinoma; GAN: Generative adversarial network; CNN: Convolutional neural network; ANN: Artificial neural network; IoU: Intersect over union.

Table 2 Applications of artificial intelligence classification methods using liver magnetic resonance imaging

Ref.	Task	Method	MR image	Accuracy	Sensitivity	Specificity	AUROC
Hectors et al[60], 2020	Stage liver fibrosis	VGG16 CNN	Gadoxetic acid-enhanced MRI	F1-4: 0.69	F1-4: 0.64	F1-4: 0.90	F1-4: 0.77
				F2-4: 0.85	F2-4: 0.82	F2-4: 0.93	F2-4: 0.91
				F3-4: 0.85	F3-4: 0.87	F3-4: 0.83	F3-4: 0.90
				F4: 0.78	F4: 0.73	F4: 0.81	F4: 0.85
Liu et al[55], 2021	Classify cHCC-CC vs non-cHCC-CC and HCC vs non-HCC	Radiomics + SVM	Gadoxetic acid-enhanced MRI	cHCC-CC vs non-cHCC-CC: 0.77	cHCC-CC vs non-cHCC-CC: 0.65	cHCC-CC vs non-cHCC-CC: 0.81	cHCC-CC vs non-cHCC-CC: 0.77
				HCC vs non-HCC: -	HCC vs non-HCC: 0.68	HCC vs non-HCC: 0.88	HCC vs non-HCC: 0.79
Wu et al[48], 2020	Classify tumours according to their LI-RADS grade	AlexNet CNN	Gadoxetic acid-enhanced MRI	0.9	1	0.835	0.95
Messaoudi et al[50], 2020	Classify tumours into HCC or non-HCC	Patch based CNN	Multiphase 3D fast spoiled gradient echo T1	0.9	?	?	?
Hamm et al[51], 2019	Classify tumours into type and LI-RADS derived classes	CNN	Multiphase contrast-enhanced T1-weighted MRI	Lesion class: 0.919	Lesion class: 0.90	Lesion class: 0.98	LI-RADS (HCC): 0.922
				LI-RADS: 0.943	LI-RADS: 0.92	LI-RADS: 0.97
Trivizakis et al[54], 2018	Classify tumours into primary or metastatic	3D CNN + SVM	Diffusion weighted MRI	0.83	0.93	0.67	0.8
He et al[65], 2019	Correctly predict liver stiffness using clinical and radiomic data	Radiomics + SVM	T2-weighted MRI	0.818	0.722	0.87	0.84
Schawkat et al[61], 2020	Stage liver fibrosis into low-stage (F0-2) and high-stage (F3-4)	Radiomics + SVM	T1-weighted MRI, T2-weighted MRI	T1-w: 0.857	?	?	T1-w: 0.82
							T2-w: 0.57
				T2-w: 0.619
Lewis et al[56], 2019	Distinguish HCC from other primary cancers	Radiomics + Binary logistic regression	Diffusion weighted MRI	Observer 1: 0.815	Observer 1: 0.793	Observer 1: 0.889	Observer 1: 0.90
				Observer 2: 0.80	Observer 2: 0.862	Observer 2: 0.778	Observer 2: 0.89
Wu et al[57], 2019	Classify tumours into HCC and HH	Radiomics + logistic regression	T2-weighted MRI, Diffusion weighted MRI, T1-weighted GRE in phase and out of phase MRI	?	0.822	0.714	0.89
Oyama et al[58], 2019	Classification of hepatic tumours into HCC, HH and MT	Radiomics + logistic regression/XGBoost	T1-weighted MRI	HCC vs MT: 0.92	HCC vs MT: 1.0	HCC vs MT: 0.84	HCC vs MT: 0.95
				HCC vs HH: 0.9	HCC vs HH: 0.96	HCC vs HH: 0.84	HCC vs HH: 0.95
				MT vs HH: 0.73	MT vs HH: 0.72	MT vs HH: 0.74	MT vs HH: 0.75
Wu et al[59], 2019	Predict pre-operative HCC grade	Combined clinical data and Radiomics + logistic regression	T2/T1-weighted	0.761	0.85	0.65	0.8
Chen et al[69], 2019	Predict pre-treatment immunscore in HCC	Combined clinical data and radiomics + multi-vote decision trees	Gadoxetic acid-enhanced MRI	0.842	0.846	0.841	0.934
Park et al[63], 2019	Stage liver fibrosis	Radiomics + logistic regression	Gadoxetic acid-enhanced MRI	F2-4: 0.803	F2-4: 0.814	F2-4: 0.784	F2-4: 0.91
				F3-4: 0.803	F3-4: 0.789	F3-4: 0.820	F3-4: 0.88
				F4: 0.813	F4: 0.921	F4: 0.754	F4: 0.87
Zhao et al[67], 2019	Predict early reoccurrence of IMCC	Combined clinical data and radiomics + logistic regression	T2-weighted MRI, gadoxetic acid-enhanced MRI	0.872	0.938	0.839	0.949
Reimer et al[68], 2018	Predict therapy response to transarterial radioembolization	Radiomics + logistic regression	Gadoxetic acid-enhanced MRI	?	Arterial phase: 0.83	Arterial phase: 0.62	Arterial phase: 0.73
					Venous phase: 0.71	Venous phase: 0.85	Venous phase: 0.76
Zhen et al[53], 2020	Classify liver tumours into benign , HCC, metastatic or other primary malignancy	CNN with clinical input	T2, diffusion, Pre- contrast T1, late arterial, portal venous, and equilibrium phase	0.919	HCC: 0.957	HCC: 0.904	HCC: 0.951
					Metastatic: 0.946	Metastatic: 1.0	Metastatic: 0.985
					Other primary: 0.733	Other primary: 0.964	Other primary: 0.989
Yasaka et al[62], 2017	Stage liver fibrosis	CNN	Gadoxetic acid-enhanced MRI	F4 vs F3-0: 0.75	F4 vs F3-0: 0.76	F4 vs F3-0: 0.76	F4 vs F3-0: 0.84
				F4-3 vs F2-0: 0.77	F4-3 vs F2-0: 0.78	F4-3 vs F2-0: 0.74	F4-3 vs F2-0: 0.84
				F4-2 vs F1-0: 0.80	F4-2 vs F1-0: 0.84	F4-2 vs F1-0: 0.65	F4-2 vs F1-0: 0.84
Kim et al[70], 2019	Predict postoperative early and late recurrence of single HCC	Radiomics + random forests	Gadoxetic acid-enhanced MRI	Harrel C-statistic: 0.716 in combined radiomic and clinicopathologic model, no significant difference to clinicopathologic model (0.696)
Kim et al[52], 2020	Detect HCC	CNN	Gadoxetic acid-enhanced MRI	0.937	0.94	0.99	0.97
Liu et al[71], 2020	Identify clinically significant portal hypertension	CNN + logistic regression	?	?	0.929	0.846	0.94

MRI: Magnetic resonance imaging; GRE: Gradient recalled echo; LI-RADS: Liver Imaging Reporting and Data System; HCC: Hepatocellular carcinoma; HH: Hepatic hemangioma; MT: Metastatic tumour; CC: Cholangiocarcinoma; cHCC-CC: Combined hepatocellular cholangiocarcinoma; IMCC: Intrahepatic mass-forming cholangiocarcinoma; CNN: Convolutional neural network; SVM: Support vector machine; AUROC: Area under the receiver operating characteristic curve.