Copyright ©The Author(s) 2021.
World J Hepatol. Dec 27, 2021; 13(12): 2039-2051
Published online Dec 27, 2021. doi: 10.4254/wjh.v13.i12.2039
Table 1 Studies applying deep learning for hepatocellular carcinoma
Data source
Deep learning
Main findings
Predicting HCC risk using clinical variables
Ioannou et al[14] 202048151 HCV cirrhosis (T: 90%, V: 10%)VHA databaseRNNClinical variablesRisk of HCC developmentRNN predicted HCC development with AUC of 0.759, and AUC of 0.806 among those who achieved SVR
Phan et al[15] 20206052 HBV and HCV (T: 70%, V: 30%)Taiwanese NHIRDCNNDisease history dataRisk of HCC developmentCNN achieved an accuracy of 0.980 and AUC of 0.886 for predicting HCC development among viral hepatitis patients
Nam et al[16] 2020T: 424 HBV cirrhosis; V: 316 HBV cirrhosis2 Korean centersResNetClinical variablesRisk of HCC developmentDL model achieved an accuracy of 0.763 and AUC of 0.782 in the validation cohort and outperformed previous models
Nam et al[17] 2020T: 349 LT recipients; V: 214 LT recipients3 Korean LT centersResNetClinical variablesRecurrent HCC after LTDL model significantly outperformed conventional models in prediction of post-T HCC recurrence with AUC of 0.75
Multi-omics-based HCC diagnosis and prognostication
Xie et al[20] 2018T: 133 HCC/54 HV; V: 52 HCC/34 HV1 center in ChinaANNGene expression HCC detectionANN using nine genes had an AUC of 0.943, 98% sensitivity, and 85% specificity for classifying HCC
Choi et al[21] 2018135 HCC (10-fold CV)TCGAG2VecGene expressionHCC prognosisG2Vec showed significantly higher prediction accuracy for patient outcomes compared to existing gene selection tools
Chaudhary et al[22] 2018T: 360 HCC; V: 220, 221, 166, 40, 27 HCCTCGA; 5 external datasetsAuto-encoderRNA-seq, miRNA-seq, methylationHCC prognosisDL model distinguished groups with survival differences and identified mutations and pathways predicting aggressive tumor behavior
Radiology-based HCC diagnosis/prediction
Streba et al[25] 2012112 FLL (10-fold CV)1 center in RomaniaANNUS imagesFLL typeANN had 87.12% testing accuracy, 93.2% sensitivity, and 89.7% specificity for classifying 5 classes of liver lesions
Hassan et al[26] 2017110 FLL (10-fold CV)1 center in EgyptAuto-encoderUS imagesFLL typeThe proposed system had 97.2% accuracy, 98% sensitivity, and 95.70% specificity for classifying liver lesions
Bharti et al[27] 201824 normal, 25 CLD, 25 cirrhosis, 20 HCC1 center in IndiaCNNUS imagesLiver stagesCNN achieved 96.6% classification accuracy for differentiating normal liver, CLD, cirrhosis, and HCC
Schmauch et al[28] 2019T: 367 FLL; V: 177 FLLCenters in FranceResNetUS imagesFLL typeDL model reached mean AUC of 0.935 for focal liver lesion detection and 0.916 for focal liver lesion characterization
Brehar et al[29] 2020T: 200 HCC; V: 68 HCC1 center in RomaniaCNNUS imagesHCC detectionCNN achieved AUC of 0.95, accuracy of 0.91, 94.4% sensitivity and 88.4% specificity for HCC detection
Jin et al[30] 2021434 HBV (3:1:1 split)1 center in ChinaDL radiomicsUS imagesRisk of HCC developmentDL radiomics model predicted 5-yr HCC development risk with AUC of 0.900 in the test set
Yasaka et al[31] 2018T: 460 liver masses; V: 100 liver masses1 center in JapanCNNCT imagesLiver mass typeCNN classified liver lesions into five categories with a median AUC of 0.92
Shi et al[32] 2020449 FLL; (T: 80%, V: 20%)1 center in ChinaCNNCT imagesFLL typeCNN applied to three-phase CT protocol images achieved AUC of 0.925 for differentiating HCC from other FLLs
Hamm et al[43] 2019T: 434 FLL; V: 60 FLL1 center in United StatesCNNMRI imagesFLL typeCNN achieved 90% sensitivity and 98% specificity for classifying FLLs and AUC of 0.992 for HCC classification
Wang et al[44] 2019T: 434 FLL; V: 60 FLL1 center in United StatesCNNMRI imagesFLL typeInterpretable DL system achieved 76.5% PPV and 82.9% sensitivity for identifying correct radiological features
Wu et al[45] 202089 liver tumors; (60: 20: 20)1 center in United StatesCNNMRI imagesLI-RADS gradingCNN achieved AUC of 0.95, 90% accuracy, 100% sensitivity and 83.5% PPV for LI-RADS grading of liver tumors
Zhen et al[46] 2020T: 1210 liver tumors; V: 201 liver tumors1 center in ChinaCNNMRI imagesLiver tumor typeCNN combined with clinical data showed AUC of 0.985 for classifying HCC with 91.9% agreement with pathology
Radiology-based HCC prognostication, treatment planning, and response to treatment
Zhang et al[47] 2021T: 158 HCC; V: 79 HCC1 center in ChinaCNNMRI imagesMVI in HCCCNN achieved AUC of 0.72, 55% sensitivity, and 81% specificity for preoperative MVI in HCC patients
Wang et al[48] 2020T: 60 HCC; V: 40 HCC1 center in ChinaCNNMRI imagesMVI in HCCFusion of deep features from MRI images yielded AUC of 0.79 for MVI prediction in HCC patients
Jiang et al[49] 2021405 HCC; (T: 80%, V: 20%)1 center in ChinaCNNCT imagesMVI in HCCCNN achieved AUC of 0.906 for prediction of MVI. Mean survival was significantly better in the group without MVI
An et al[50] 2020141 single HCC resect MWA1 center in ChinaCNNMRI imagesAblative marginDeep learning model accurately estimated ablative margins and risk of local tumor progression
Liu et al[51] 2020T: 89 HCC resect TACE; V: 41 HCC rec. TACE1 center in ChinaCNNUltrasound imagesResponse to TACEDeep learning radiomics model predicted tumor response to TACE with AUC of 0.93
Peng et al[52] 2020T: 562 HCC resect TACE; V:227 HCC rec. TACE3 centers in ChinaCNNCT imagesResponse to TACEDeep learning model had accuracies of 85.1% and 82.8% for predicting TACE response in 2 validation cohorts
Liu et al[53] 2020243 HCC resect TACE (6:1:3 split)1 center in ChinaCNNCT imagesPost-TACE survivalHigher DL score was an independent prognostic factor and predicted overall survival with AUCs of 0.85-0.90
Zhang et al[54] 2020201 HCC resect TACE + sorafenib (T: 120, V: 81)3 centers in ChinaCNNCT imagesOS on TACE + sorafenibDeep learning signature achieved C-index of 0.714 for predicting OS in HCC patients receiving TACE + sorafenib
Histopathology-based HCC diagnosis, subtyping, and outcome predictions
Lin et al[55]2019113 HCC1 center in ChinaCNNHistopath imagesHCC differentiationCNN achieved an accuracy of 0.941 for determining HCC differentiation on multiphoton microscopy
Kiani et al[56] 202070 WSI (35 HCC, 35 CC)TCGACNNHistopath imagesHCC vs CCCNN-based “Liver Cancer Assistant” accurately differentiated HCC vs cholangiocarcinoma
Liao et al[57] 2020T: 491 HCC; V: 455 HCCTCGA; 1 center in ChinaCNNHistopath imagesHCC detection, mutationsCNN distinguished HCC from adjacent tissues with AUC of 1.00 and predicted specific mutations with AUC over 0.70
Wang et al[58] 2020T: 99 HCC; V: 205 HCCTCGACNNHistopath imagesHistological HCC subtypeUnsupervised clustering identified 3 histological subtypes complementing molecular pathways and prognostic value
Chen et al[59] 2020T: 402 HCC/89 normal; V: 67 HCC/34 normalGDC portal; 1 center in ChinaCNNHistopath imagesHCC grade mutationsCNN achieved 89.6% accuracy for tumor differentiation stage and predicted presence of specific gene mutations
Lu et al[60] 2020421 HCC/105 normal (6-fold CV)GDC portalCNNHistopath imagesHCC prognosisPre-trained CNN predicted OS using pathology images and identified HCC subgroups with different prognosis
Saillard et al[61] 2020T: 194 HCC; V: 328 HCC1 French center TCGACNNHistopath imagesSurvival after HCC resectionCNN models using pathology images predicted survival with C-index 0.75-0.78 and outperformed conventional models
Shi et al[62] 2021T: 1125 HCC; V: 320 HCC1 center in China; TCGACNNHistopath imagesHCC outcomesDeep learning-based “tumor risk score” was superior to clinical staging and stratified 5 groups of different prognosis
Yamashita et al[63] 2021T: 36 WSI; V: 30 WSI1 center in United States; TCGACNNHistopath imagesPost-surgical recurrenceCNN risk scores outperformed TNM system for predicting recurrence and identified high-and low-risk subgroups