Radiomics in hepatocellular carcinoma: A state-of-the-art review

Table 1 Some representative studies of radiomics in hepatocellular carcinoma

Ref.	Application task	Study design	Imaging modality	Radiomics features	Algorithm	Sample size	Training set	Test/validation set	Performance
Liu et al[29], 2021	Differentiation of cHCC-CC from HCC and CC	Retrospective, single-center	CT, MRI	1419	SVM	85 patients with HCC (37), cHCC-CC (24) and CC (24)	85	NA	Excellent performance for differentiation of HCC from non-HCC (AUC: 0.79-0.81 in MRI, AUC: 0.71-0.81 in CT)
Nie et al[32], 2020	Differentiation of HCA from HCC	Retrospective, two-institutes	CT	3768	mRMR, LASSO	131 patients with HCC (85) and HCA (46)	93	38	Favorable performance (AUC: 0.96 in training set, AUC: 0.94 in test set)
Wu et al[33], 2019	Pathological grade of HCC	Retrospective, single-center	MRI	656	LASSO	170 patients with HCCs	125	45	Radiomics signature model outperformed the clinical factors-based model; the combined model achieved the best performance (AUC: 0.80)
Mao et al[38], 2020	Pathological grade of HCC	Retrospective, single-center	CT	3376	RFE, XGBoost	297 patients with HCCs	237	60	The radiomics signatures combined with clinical factors significantly achieved the best performance (AUC: 0.8014)
Xu et al[43], 2019	Preoperative prediction of MVI in HCC	Retrospective, single-center	CT	7260	Ref-SVM, Multivariable logistic regression	495 patients with HCC	300	145 (test); 50 (validation)	Good performance (AUC: 0.909 in the training/validation set, AUC: 0.889 in the test set)
Chong et al[47], 2021	Preoperative prediction of MVI in HCC	Retrospective, single-center	MRI	854	LASSO, RF, logistic regression	356 patients with HCCs ≤ 5 cm	250	106	AUC: 0.920 using RF; AUC: 0.879 using logistic regression (in validation set)
Fu et al[54], 2019	Assistant in optimal treatment choices of HCC between LR and TACE	Retrospective, multi-center (5 institutions)	MRI	708	LASSO, Akaike information criterion	520 patients with HCC	302	218	Good discrimination and calibrations for 3-year PFS (AUC: 0.80 in training set, AUC: 0.75 in validation set); threshold ≤ -5.00: suggesting LR, threshold > -5.00: suggesting TACE
Sun et al[56], 2020	Predicting the outcome of TACE for unresectable HCC	Retrospective, single-center	MRI	3376	LASSO, multivariable logistic regression	84 patients with BCLC B stage HCC	67	17	The radiomics signatures combined with clinical factors significantly achieved the best performance (AUC: 0.8014)
Ji et al[66], 2020	Predicting early recurrence after LR	Retrospective, multi-center (3 institutions)	CT	846	LASSO-Cox regression	295 patients with HCC	177 (Institution 1)	118 (Institution 2 and 3, external validation)	Better prognostic ability (C-index: 0.77, P < 0.05), lower prediction error (integrated brier score: 0.14), and better clinical usefulness than rival models and staging systems
Zhao et al[67], 2020	Predicting early recurrence after LR	Retrospective, single-center	MRI	1146	LASSO, stepwise and multivariable logistic regression	113 patients with HCC	78	35	The nomogram integrating the Rad score and clinicopathologic-radiologic risk factors showed better discrimination and clinical utility (AUC: 0.873)
Wang et al[75], 2020	Predicting 5-year survival after LR	Retrospective, multi-center (2 institutions)	MRI	3144	RF, multivariate logistic regression	201 patients with HCC	160	41 (five-fold cross-validation)	The model incorporating the radiomics signature and clinical risk factors obtained good calibration and satisfactory discrimination (AUC: 0.9804 in training set, AUC: 0.7578 in validation set)
Song et al[76], 2020	Predicting RFS after TACE	Retrospective, single-center	MRI	396	LASSO-Cox regression, multivariate Cox regression	184 patients with HCC	110	74	The model using the radiomics signature with the clinical-radiological risk factors showed the best performance (C-index: 0.802)

cHCC-CC: Combined hepatocellular cholangiocarcinoma; NA: Not available; HCC: Hepatocellular carcinoma; CC: Cholangiocarcinoma; CT: Computed tomography; MRI: Magnetic resonance imaging; GLCM: Gray-level co-occurrence matrix; SVM: Support vector machine; AUC: Area under the receiver operating characteristic curve; HCA: Hepatic adenoma; mRMR: Maximal relevance and minimum redundancy; LASSO: Least absolute shrinkage and the selection operator; RFE: Recursive feature elimination; XGBoost: eXtreme gradient boosting; MVI: Microvascular invasion; Ref-SVM: Recursive feature selection support vector machine; RF: Random forest; LR: Liver resection; TACE: Transarterial chemoembolization; PFS: Progression-free survival; BCLC: Barcelona clinic liver cancer; C-index: Concordance index; RFS: Recurrence free survival.

Table 2 Some representative studies of deep learning in hepatocellular carcinoma

Ref.	Application task	Study design	Imaging modality	Algorithm	Sample size	Training set	Test/validation set	Performance
Bousabarah et al[91], 2021	Automatic detection and segmentation of HCC	Retrospective, single-center	CT	DCNN, U-net	174 patients with 231 lesions	165	33 (test); 33 (validation)	Mean DSC between automatically detected lesions using the DCNN + RF + TR and corresponding manual segmentations: 0.64/0.68 (validation/test), and 0.91/0.91 for liver segmentations
Yasaka et al[92], 2018	Differentiation of HCC and other liver tumors	Retrospective, single-center	CT	CNN	560 patients	460	100	Accuracy: 84% in test set
Hamm et al[93], 2019	Diagnosis and classification of HCC	Retrospective, single-center	MRI	CNN	494 patients	434	60	Accuracy: 92%, AUC: 0.992
Yamashita et al[95], 2020	Diagnosis and categorization of HCC with LI-RADS	Retrospective, multi-center	CT, MRI	CNN	314 patients (163 CT, 151 MRI)	220	47 (test); 47 (internal validation); 112 (external validation)	Overall accuracy: 60.4% and AUCs: 0.85, 0.90, 0.63, and 0.82 for LR-1/2, LR-3, LR-4, and LR-5, respectively
Wang et al[96], 2020	Preoperative prediction of MVI in HCC	Retrospective, single-center	MRI	CNN	97 patients with 100 HCCs	60 HCCs	40 HCCs	The combination of deep features from the b = 0, b = 100, b = 600, and ADC images presented the best results (AUC: 0.79)
Peng et al[97], 2020	Prediction of treatment response of TACE	Retrospective, multi-center (3 institutions)	CT	ResNet50	789 patients with HCC	562 (Institution 1)	89 (Institution 2); 138 (Institution 3)	Excellent predictive performance for CR, PR, SD, and PD (accuracy: 84.3%; AUCs: 0.97, 0.96, 0.95, and 0.96 in training set, accuracies: 85.1% and 82.8% in the two validation sets)
Zhang et al[98], 2020	Predicting OS after TACE + Sorafenib	Retrospective, multi-center (3 institutions)	CT	DenseNet (CNN)	201 patients with HCC	120 (Institutions 1 and 2)	81 (Institution 3)	Favorable prediction performance (C-index: 0.717 in training set, C-index: 0.714 in validation set)
Tamada et al[99], 2020	Motion artifact reduction	Retrospective, single-center	MRI	CNN	34 patients with HCC	14	20	Significant reduction of the magnitude of the artifacts and blurring induced by respiratory motion
Esses et al[100], 2018	Automated image quality evaluation	Retrospective, single-center	MRI	CNN	522 patients with HCC	351	171	High negative predictive value (94% and 86% relative to two readers)

DCNN: Deep convolutional neural network; TR: Thresholding; DSC: Dice similarity coefficient; CNN: Convolutional neural network; LI-RADS: Liver imaging reporting and data system; CR: Complete response; PR: Partial response; SD: Stable disease; PD: Progressive disease.