Introduction
Methods
Results
Study | Group | Type of paper | Type of journal | Type of ML method used | Test Set | External validity | ML to clinician | Public dataset |
---|---|---|---|---|---|---|---|---|
Mubashir et al 2019 [19] | Liver segmentation | A | Non-medical | DBN-DNN | 15 + 15 (2 open datasets) | No | No | Yes |
Mubashir et al 2019 [19] | Liver segmentation | PP | Non-medical | CNN | 5 | No | No | Yes |
Ahn et al 2020 [20] | Liver segmentation | A | Medical | 3d U-net, DLA DeepLabV3 | 20 CT series and 60 CT series | Yes | Yes | Yes |
Bhavya et al 2018 [21] | Liver segmentation | PP | Medical | Real AdaBoost classifier | 70 | Yes | No | Yes |
Albishri 2019 [22] | Liver segmentation | PP | Non-medical | Cascade U-net | 32 patient’s data (unclear about the total number of data) | No | No | No |
Ali 2017 [23] | Liver segmentation | PP | Non-medical | SVM | 50 | No | Yes | Yes |
Alirr 2020 [24] | Liver segmentation | A | Non-medical | U NET + level set | 20 | Yes | Yes | Yes |
Astono 2018 [25] | Liver segmentation | PP | Non-medical | CNN-adjacent net | 10 scans | Yes | No | Yes |
Ben-Cohen 2016 [26] | Liver segmentation | PP | Non-medical | FCN-VGG 16-layer net | 70 CT sets | Yes | No | Yes |
Bevilacqua et al 2017 [27] | Liver segmentation | PP | Non-medical | ANN classifier by using mono-objective genetic algorithm (GA) | Not mentioned | Yes | No reference | No |
Bhole 2011 [28] | Liver segmentation | PP | Non-medical | MRF | 10 series of 10 patients | No | No | Yes |
Budak et al 2020 [29] | Liver segmentation | A | Medical | CEDCNN | 5 sets (589 slices) | No | No | Yes |
Cai 2019 [30] | Liver segmentation | A | Medical | Adaptive scale-kernel fuzzy clustering models | They have created 3 model from different dataset, and the fourth model for fine tuning. Difficult to give in number of patients used in for training. As they have used transfer learning from one model to another model where adding some more data | Yes | No | Yes |
Chen 2019 [31] | Liver segmentation | PP | Non-medical | MPNet, adversarial densely connected network and a deep FCNN | 10 | Yes | No | Yes |
Chen et al 2019 [31] | Liver segmentation | A | Medical | Channel-U-net, a spatial channel wise convolutional neural network | Yes | No | Yes | |
Chlebus 2018 [32] | Liver segmentation | A | Non-medical | FCNN and object based postprocessing | Not mentioned | Yes | Yes | No |
Choi et al 2018 [33] | Liver segmentation | A | Medical | CNN | 150 images | Yes | Yes | No |
Chung 2020 [34] | Liver segmentation | A | Non-medical | CENet | 28 volumes | No | No | Yes |
Danciu 2013 [35] | Liver segmentation | A | Medical | SVM | 76 patients (20–40 images of the liver per patient) | No | Yes | Yes |
Danciu 2012 [36] | Liver segmentation | PP | Non-medical | 3D DCT and SVM | 26,608 images of 70 CT scans from 40 patients | No | No | Yes |
Delmoral 2019 [37] | Liver segmentation | PP | Non-medical | CNN | 31CT | Yes | No | Yes |
Dong 2020 [38] | Liver segmentation | A | Non-medical | HDCNN | 50 patients, 1272 images | No | No | Yes |
Dou et al 2016 [39] | Liver segmentation | A | Medical | 3D deeply supervised network | 5 patients for testing, 5 patients for validation | Yes | No | No |
Guo 2019 [40] | Liver segmentation | A | Medical | FCNN | – | Yes | No reference | No |
He et al 2016 [41] | Liver segmentation | A | Medical | Ada Boost guided active shape model | (1) 46 lesions for validation, 46 lesions for testing; (2) not specified | Yes | No | Yes |
Heker 2019 [42] | Liver segmentation | PP | Non-medical | Cascade U-net | Not specified | Yes | No | No |
Hu 2016 [43] | Liver segmentation | A | Medical | 3D-CNN | 10 patients | Yes | No | Yes |
Huang et al 2012 [44] | Liver segmentation | PP | Non-medical | ELM | Not specified | Yes | No | No |
Ji 2013 [45] | Liver segmentation | A | Non-medical | ACM | Not specified | Yes | No | No |
Jiang 2018 [46] | Liver segmentation | A | Medical | Registration based organ positioning, FCMC, ELM, ACM | Not specified | No | No | Yes |
Jiang 2019 [47] | Liver segmentation | A | Non-medical | 3D FCN, AHCBlocks | 12 images for validation, 1 for testing | Yes | No | Yes |
Jin 2017 [48] | Liver segmentation | PP | Non-medical | FCN-U-net | 25 patients for testing, 25 patients for validation | Yes | No | Yes |
Kavur et al 2020 [49] | Liver segmentation | A | Medical | CNN | 20 patients | No | No | Yes |
Kumar 2016 [50] | Liver segmentation | A | Medical | Feedforward neural network | Not mentioned | No | No reference | No |
Chung 2020 [34] | Liver segmentation | A | Non-medical | CNN (CENet) | 150 images | No | No | No |
Zheng et al 2019 [51] | Liver segmentation | PP | Non-medical | GAN + deep atlas prior | 28 volumes | No | No | Yes |
Zhang, Y. 2018 [52] | Liver segmentation | PP | Non-medical | FCN + CRF | 5 scans/patients | No | No | Yes |
Zhang, L. 2018 [53] | Liver segmentation | PP | Non-medical | U-net | 76 patients (20–40 images of the liver per patient) | No | No | Yes |
Xu 2019 [54] | Liver segmentation | PP | Non-medical | RES-U-Net, connected components analyzing and CRF | 26,608 images of 70 CT scans from 40 patients | No | No | Yes |
Xi 2020 [55] | Liver segmentation | A | Non-medical | Cascade U-RES-Net (CNN + Dlo + TL + GDL + GTL) | 70 image sets | No | No | Yes |
Xin 2020 [56] | Liver segmentation | A | Non-medical | CNN | 32 patients, 643 slices containing lesions | No | No | No |
Xia 2019 [57] | Liver segmentation | A | Non-medical | CNN Deep Adversarial Networks (DeepLab-v3) + weighted loss function | 8800 images | No | No | Yes |
Winkel et al 2020 [58] | Liver segmentation | A | Medical | DRL (CNN + RL) | 20 sets, 6 sets per patient | No | Yes | No |
Wang et al 2019 [59] | Liver segmentation | PP | Non-medical | CNN | 28 patients | No | No | Yes |
Tian 2019 [60] | Liver segmentation | PP | Non-medical | U-net (GLC-U-net, CNN) | 50 patients, 1272 images | No | No | Yes |
Tang 2017 [61] | Liver segmentation | PP | Non-medical | FCN (+ level set) | 5 patients for testing, 5 patients for validation | No | No | No |
Seo et al 2020 [62] | Liver segmentation | A | Non-medical | CNN (modified U-Net) | (1) validation: 5 patients; 2550 images; testing: 35 patients; 16,125 images; (2) 5 patients, 525 images | Yes | No | Yes |
Selvi 2014 [63] | Liver segmentation | PP | Non-medical | High-order neural network | Not specified | No | No | No |
Selvathi et al 2013 [64] | Liver segmentation | PP | Non-medical | ELM + FCMC | Not specified | No | No | No |
Sayed 2016 [65] | Liver segmentation | PP | Non-medical | Fuzzy clustering + GWO (Liver and liver lesion segmentation); SVM (liver disease classification: benign/malignant) | Not provided | No | No | No |
Sakboonyara 2019 [66] | Liver segmentation | PP | Non-medical | U-Net, 2D (CNN/ FCN) | 5 images | No | No | Yes |
K S et al 2018 [67] | Liver segmentation | PP | Non-medical | U-Net and 3D CRF | – | No | No | No |
Raj 2016 [68] | Liver segmentation | PP | Non-medical | SVM | Not specified | No | No | No |
Rafiei 2018 [69] | Liver segmentation | PP | Non-medical | FCN + CRF | 10 patients | No | No | Yes |
Qin et al 2018 [70] | Liver segmentation | A | Medical | CNN (SBBS-CNN, based on CifarNet) | Not specified | No | No | Yes |
Ponnoprat et al 2020 [71] | Liver segmentation | A | Non-medical | U-Net for segmentation + CRF for post-processing + SVM for classification | 17 patients, 2042 images | No | No | No |
Ouhmich 2019 [72] | Liver segmentation | A | Medical | U-Net | Not specified | No | No | No |
Ng et al 2020 [73] | Liver segmentation | A | Medical | Gaussian mixture model and U-Net | 6 patients (fivefold cross validation) | No | No | No |
Nayak et al 2019 [74] | Liver segmentation | A | Medical | Segmentation: region-growing; classification: SVM | Not specified | No | No | Yes |
Mukherjee et al 2013 [75] | Liver segmentation | PP | Non-medical | SVM + PCA | Not specified | No | No | No |
Morshid et al 2019 [76] | Liver segmentation | A | Medical | Segmentation: U-Net, 2D; prediction: RFC | Not specified | Yes | Yes | Yes |
Mohagheghi and Foruzan 2020 [77] | Liver segmentation | A | Medical | U-Net | 12 images for validation, 1 for testing | No | No | Yes |
Mofrad 2014 [78] | Liver segmentation | A | Medical | Classification: SVM, k-NN | 1 patient | No | No | No |
Meng L 2020 [79] | Liver segmentation | A | Non-medical | TDP-CNN + CRF (post-processing) | 25 patients for testing, 25 patients for validation | No | No | Yes |
Luo and Li 2014 [80] | Liver segmentation | PP | Non-medical | SVM | 1 image, 1 patient | No | No | Yes |
Lu et al 2017 [81] | Liver segmentation | A | Medical | CNN + graph cut | SLiver07: 10 patients; 3D-IRCADb: 20 patients | Yes | Yes | Yes |
Selvaraj 2013 [82] | Liver segmentation | PP | Non-medical | Lesion segmentation: FCM; feature selection: BPSO; classification: PNN | 15 images | No | No | No |
Li 2014 [83] | Liver segmentation | PP | Non-medical | PCA + ASM + k-NN | 5 whole body scans, 5 abdominal contrast-enhanced scans | No | No | Yes |
Li et al 2018 [84] | Liver segmentation | A | Non-medical | H-Dense U-Net | LiTS 2017: 70 patients; 3D-IRCADb: cross-validation | Yes | No | Yes |
Liu et al 2019 [85] | Liver segmentation | A | Non-medical | U-Net + graph cut | 20 patients | No | No | Yes |
Linguraru et al 2012 [86] | Liver segmentation | A | Non-medical | SVM | LiTS 2008: 4 patients; SLiver07: 10 patients | Yes | No | Yes |
Astono et al 2018 [25] | Liver segmentation | A | Non-medical | Adjacent Net | Validation: 2 patients; test: 2 × 10 patients | No | No | Yes |
Afifi and Nakaguchi 2015 [87] | Liver segmentation | PP | Non-medical | MSCA + graph cut in detection | Not specified | No | No reference | No |
Roth 2020 [88] | Liver segmentation | PP | Non-medical | U-net | 70 | Yes | No | Yes |
Tran 2021 [89] | Liver segmentation | A | Non-medical | U-Net multilayer | 30 scan (15 CT from each datasets) | Yes | No | Yes |
Xu et al 2020 [90] | Liver segmentation | PP | Non-medical | pyramidal U-net | fourfold cross-validation | No | No reference | 0 |
Yu et al 2021 [91] | Liver segmentation | A | Non-medical | DResU-Net | 25 | Yes | No | Yes |
Zhang, Y et al 2021 [92] | Liver segmentation | A | Non-medical | RECIST NET | 46 | No | No | 0 |
Zhang, Yao 2021 [92] | Liver segmentation | PP | Non-medical | CNN (deep attentive refinement network) | 70 | Yes | No | Yes |
Ayalew 2021 [93] | Liver segmentation | A | Non-medical | U-net | 392 images | No | No | Yes |
Chen et al 2020 [94] | Liver segmentation | A | Medical | U-net | 300 images | Yes | No | Yes |
Chung 2021 [95] | Liver segmentation | A | Medical | CNN | 80 patients | No | No | Yes |
Elmenabawy et al 2020 [96] | Liver segmentation | PP | Non-medical | CDNN | 33 patients | No | No | Yes |
Fan 2020 [97] | Liver segmentation | A | Non-medical | U-net multi-scale attention net | 70 patients | No | No | Yes |
He et al 2021 [98] | Liver segmentation | A | Medical | U-net (3D RA-U-Net) | 252 images, 63 patients | Yes | Yes | Yes |
Kwon 2020 [99] | Liver segmentation | PP | Non-medical | U-net | 70 patients | No | No | Yes |
Lei 2020 [100] | Liver segmentation | PP | Non-medical | U-Net / V-Net | 31 patients | No | No | Yes |
Afifi 2015 [87] | Lesion detection | PP | Non-medical | Mean-shift segmentation algorithm | 15 patients 169 lesions | No | No | No |
Ali et al 2017 [23] | Lesion detection | PP | Non-medical | SVM | 50 | No | Yes | Yes |
Ben-Cohen 2016 [26] | Lesion detection | PP | Non-medical | FCN-VGG 16 layer net | 70 CT sets | Yes | No | Yes |
Ben-Cohen 2018 [101] | Lesion detection | A | Non-medical | FCN8 net-VGG 16 layer net, and sparsity-based dictionary learning (localized patch level analysis usin superpixel sparse based classification | 7 data sets | No | No | Yes |
Bevilacqua et al 2017 [102] | Lesion detection | PP | Non-medical | ANN clssifier by using mono-objective GA | Not mentioned | Yes | No reference | No |
Bevilacqua et al 2017 [102] | Lesion detection | PP | Non-medical | ANN clssifier by using MOGA | Not mentioned | No | No reference | No |
Chen et al 2019 [103] | Lesion detection | PP | Non-medical | Dual-attention dilated residual network-weakly supervised localization | 10 + 10 dataset from Sliver | No | No | Yes |
Frid-Adar 2017 [104] | Lesion detection | PP | Non-medical | Multi-class patch based CNN system | (1) Validation: 5 patients, 2550 images, testing: 35 patients, 16,125 images; (2) 5 patients, 525 images | No | No | Yes |
Furuzuki et al 2019 [105] | Lesion detection | PP | Non-medical | Faster R-CNN | Not specified | No | No | No |
Gong et al 2019 [106] | Lesion detection | A | Medical | R-CNN, partial least square regression discriminant analysis model | 5images | No | Yes | Yes |
Huang et al 2013 [107] | Lesion detection | PP | Non-medical | Kernel-based ELM with classifier | 17 patients, 2042 images | No | No | No |
Jin 2017 [108] | Lesion detection | PP | Non-medical | CNN + ensemble learning | SLiver07: 10 patients; 3D-IRCADb: 20 patients | No | No | Yes |
Jin 2015 [109] | Lesion detection | PP | Non-medical | Improved back propagation neural network | 15 images | No | No | No |
Kim 2019 [110] | Lesion detection | PP | Non-medical | Cycle-Consistent CNN | Not specified | No | No | No |
Vivanti 2017 [111] | Lesion detection | A | Medical | CNN + RFC | Not specified | No | No | No |
Tao et al 2019 [112] | Lesion detection | PP | Non-medical | FCN + RPN | ∼ 5000 images for testing and ∼ 5000 images for validation | No | No | Yes |
Liang et al 2019 [113] | Lesion detection | PP | Non-medical | CNN (recurrant with long short-term memory) | (1) validation: 175; test: 153; (2) validation: 175; test: 153 | No | No | No |
Lee 2018 [114] | Lesion detection | PP | Non-medical | SSD | fivefold cross-validation | No | No | No |
Afifi 2015 [87] | Lesion detection | PP | Non-medical | MSCA (+ graph cut in detection) | Not specified | No | No reference | No |
Yang et al 2021 [115] | Lesion detection | A | Non-medical | CNN | 337 | Yes | No reference | 0 |
Zhou et al 2021 [116] | Lesion detection | A | Medical | CNN | 1/4 of lesion was used for testset | No | No | 0 |
Albishri 2019 [22] | Lesion segmentation | PP | Non-medical | Cascade U-net | 32 patients data (unclear about the total number of data) | No | No | No |
Alirr 2020 [24] | Lesion segmentation | A | Non-medical | U NET + level set | 20 | Yes | Yes | Yes |
Almotairi 2020 [117] | Lesion segmentation | A | Non-medical | Modified Seg Net | 20 CT from local hospital | No | No | Yes |
Anter 2019 [118] | Lesion segmentation | A | Medical | Fast fuzzy C-means and adaptive watershed algorithm | 30 | Yes | No | Yes |
Budak 2020 [29] | Lesion segmentation | A | Medical | CEDCNN | 5 sets (589 slices) | No | No | Yes |
Chen, L. 2019 [103] | Lesion segmentation | PP | Non-medical | MPNet, adversarial densely connected network and a deep FCNN | 10 | Yes | No | Yes |
Chen, X. et al 2019 [119] | Lesion segmentation | PP | Non-medical | FED-Net | 10 CT series | No | No | Yes |
Chen, Y. et al 2019 [31] | Lesion segmentation | A | Medical | Channel-U -net | Yes | No | Yes | |
Chlebus 2018 [32] | Lesion segmentation | A | Non-medical | FCNN- and object-based postprocessing | Not mentioned | Yes | Yes | No |
Delmoral 2019 [37] | Lesion segmentation | PP | Non-medical | CNN | 31CT | Yes | No | Yes |
Deng 2019 [120] | Lesion segmentation | A | Medical | Dynamic regulation to functional parameters over iterations using the 3D CNN | 20 sets, 6 sets per patient | Yes | No | No |
Dong 2020 [38] | Lesion segmentation | A | Non-medical | HDCNN | 50 patients, 1272 images | No | No | Yes |
Heker 2019 [42] | Lesion segmentation | PP | Non-medical | Cascade U-net | Not specified | Yes | No | No |
Huang et al 2013 [107] | Lesion segmentation | PP | Non-medical | Kernel-based ELM with classifier | 17 patients, 2042 images | No | No | No |
Huang et al 2014 [121] | Lesion segmentation | PP | Non-medical | Random feature subspace ensemble–based ELM | 6 patients (fivefold cross-validation) | No | No | No |
Jiang 2018 [46] | Lesion segmentation | A | Medical | Registration based organ positioning, fuzzy C means clustering and ELM, ACM | Not specified | No | No | Yes |
Jiang 2019 [47] | Lesion segmentation | A | Non-medical | 3D FCN composed of multiple AHCBlocks | 12 images for validation, 1 for testing | Yes | No | Yes |
Kadoury 2015 [122] | Lesion segmentation | A | Medical | Grassmanian kernels and discriminant manifold, CRF | 5 whole body scans, 5 abdominal contrast-enhanced scans | Yes | No | Yes |
Almotairi 2020 [117] | Lesion segmentation | A | Non-medical | Modified SegNet | 3 patients, 454 images for testing and 45 for validation | No | No | Yes |
Zhou 2013 [123] | Lesion segmentation | PP | Non-medical | CNN | 16 patients | Yes | No | No |
Zhang, Yue et al 2020 [124] | Lesion segmentation | A | Non-medical | 2D U-net + 3D FCN and unsupervised fuzzy c-means clustering | (1) 36 images; (2) 70 images | Yes | No | Yes |
Zhang, Yi 2020 [125] | Lesion segmentation | A | Non-medical | CNN | (1) 9 sets for testing, 20 for verification/validation; (2) 5 for testing, 5 for verification | Yes | No | Yes |
Zhang, Xing 2011 [126] | Lesion segmentation | PP | Non-medical | SVM + traditional feature extraction | Not specified | No | No | Yes |
Xi 2020 [55] | Lesion segmentation | A | Non-medical | Cascade U-RES-Net (CNN + Dlo + TL + GDL + GTL) | 70 image sets | No | No | Yes |
Xin 2020 [56] | Lesion segmentation | A | Non-medical | CNN | 32 patients, 643 slices containing lesions | No | No | No |
Wu 2019 [127] | Lesion segmentation | PP | Non-medical | MW-U-net | 15 patients, 100–135 images per patient | No | No | Yes |
Wei et al 2019 [128] | Lesion segmentation | PP | Non-medical | CNN (HMMMNet) | (1) LiTS 2017: 26 patients; (2) decathlon: 70 (not specified in the article—found at medicaldecathlon.com) | No | No | Yes |
Vorontsov et al 2018 [129] | Lesion segmentation | PP | Non-medical | CNN (FCN) | 15 patients | No | No | Yes |
Vorontsov et al 2017 [130] | Lesion segmentation | A | Non-medical | MLP | 5 patients | No | No | No |
Vivanti 2017 [111] | Lesion segmentation | A | Medical | CNN + RFC | Not specified | No | No | No |
Vivanti 2018 [129] | Lesion segmentation | A | Non-medical | CNN (× 2: global and individual) | Not specified | No | No | No |
Todoroki 2019 [131] | Lesion segmentation | PP | Non-medical | CNN | 266,000, 282,000, and 215,000 patch images (tested once each) | No | No | No |
Sun 2017 [132] | Lesion segmentation | PP | Non-medical | FCN | (1) 3D-IRCADb: 40 images; (2) JDRD: 36 images | Yes | No | Yes |
Shimizu 2013 [133] | Lesion segmentation | A | Non-medical | U-Boost | Not specified | No | No | No |
Seo 2020 [62] | Lesion segmentation | A | Non-medical | Modified U-Net | (1) Validation: 5 patients; 2550 images; testing: 35 patients, 16,125 images; (2) 5 patients, 525 images | Yes | No | Yes |
Selvathi et al 2013 [64] | Lesion segmentation | PP | Non-medical | ELM + FCMC | Not specified | No | No | No |
Sayed 2016 [65] | Lesion segmentation | PP | Non-medical | Segmentation: fuzzy clustering + GWO; classification: SVM | Not provided | No | No | No |
Raj 2016 [68] | Lesion segmentation | PP | Non-medical | SVM | Not specified | No | No | No |
Ouhmich 2019 [72] | Lesion segmentation | A | Medical | U-Net | Not specified | No | No | No |
Morshid 2019 [76] | Lesion segmentation | A | Medical | Segmentation: U-Net; prediction: RFC | Not specified | Yes | Yes | Yes |
Moawad et al 2020 [134] | Lesion segmentation | A | Medical | U-Net | Not specified | No | Yes | No |
Meng et al 2020 [79] | Lesion segmentation | A | Non-medical | TDP-CNN + CRF | 25 patients for testing, 25 patients for validation | No | No | Yes |
Selvaraj 2013 [82] | Lesion segmentation | PP | Non-medical | Segmentaion: FCM; feature selection: BPSO; classification: PNN | 15 images | No | No | No |
Li et al 2018 [84] | Lesion segmentation | A | Non-medical | H-DenseU-Net | LiTS 2017: 70 patients; 3D-IRCADb: cross-validation | Yes | No | Yes |
Radu et al 2020 [135] | Lesion segmentation | A | Medical | CNN | 30 CT for testing | Internal | No | 0 |
Roth 2020 [88] | Lesion segmentation | PP | Non-medical | U-net | 70 | External | No | Yes |
Xin 2020 [56] | Lesion segmentation | A | Medical | CNN | 643 slice for test | No | No | 0 |
Tran 2021 [89] | Lesion segmentation | A | Non-medical | U-Net multilayer | 30 scan (15 ct from each datasets) | Yes | No | Yes |
Haq et al 2021 [136] | Lesion segmentation | PP | Non-medical | Resnet R-CNN | 70 | Yes | No | Yes |
Yang et al 2021 [115] | Lesion segmentation | A | Non-medical | CNN | 337 | Yes | No reference | 0 |
Yu et al 2021 [91] | Lesion segmentation | A | Non-medical | DResU-Net | 25 | Yes | No | Yes |
Zhang, Yao 2021 [92] | Lesion segmentation | PP | Non-medical | CNN (deep attentive refinement network) | 70 | Yes | No | Yes |
Anil 2021 [137] | Lesion segmentation | A | Non-medical | MDCN + FRN | NA | No | No | Yes |
Aslam et al 2021 [138] | Lesion segmentation | A | Non-medical | ResU-Net | NA | No | No | Yes |
Ayalew 2021 [93] | Lesion segmentation | A | Non-medical | U-net | 392 images | No | No | Yes |
Chen et al 2021 [94] | Lesion segmentation | A | Medical | U-net | 300 images | Yes | No | Yes |
Dey 2020 [139] | Lesion segmentation | PP | Non-medical | CNN | 70 patients | No | No | Yes |
Elmenabawy et al 2020 [96] | Lesion segmentation | PP | Non-medical | CDNN (conv-deconv neural net) | 33 patients | No | No | Yes |
Fan 2020 [97] | Lesion segmentation | A | Non-medical | U-net (multi-scale attention net) | 70 patients | No | No | Yes |
Hamard et al 2020 [140] | Lesion segmentation | A | Medical | NA (off the shelf product) | 44 | Yes | Yes | No |
He et al 2021 [98] | Lesion segmentation | A | Medical | U-net (3D RA-U-Net) | 252 images, 63 patients | Yes | Yes | Yes |
Kwon 2020 [99] | Lesion segmentation | PP | Non-medical | U-net | 70 patients | No | No | Yes |
Adcock 2014 [18] | Classification | A | Non-medical | SVM-LibSVM (multidimensional scaling (CMDS) | Not mentioned | No | No | No |
AmirHosseini 2019 [141] | Classification | A | Non-medical | Fuzzy inference system | 7 patients for HCC segmentation, 20 patients for liver segmentation | No | No | Yes |
Balagourouchetty et al 2020 [142] | Classification | A | Non-medical | GoogLeNet based Ensemble FCNet Classifier | Not mentioned exactly number but they have 10% data to test set and have used tenfold cross-validation | No | No | Yes |
Bevilacqua et al 2017 [27] | Classification | PP | Non-medical | ANN classifier by using mono-objective genetic algorithm (GA) | Not mentioned | Yes | No reference | No |
Cao et al 2020 [143] | Classification | A | Medical | Multiphase convolutional dense network | 42CT (12 from local and 20 + 10 from Sliver07) | No | Yes | Yes |
Chen et al 2019 [103] | Classification | PP | Non-medical | Dual-attention dilated residual network—weakly supervised localization | 10 + 10 dataset from Sliver | No | No | Yes |
Das 2019 [144] | Classification | A | Medical | Watershed Gaussian–based deep learning, DNN | 32 patients, 643 slices containing lesions | No | No | No |
Devi 2020 [145] | Classification | A | Non-medical | Region growing process for liver segmentation = > kernalized fuzzy C-means algorithm for lesion extraction, SVM-based classifier for classification of tumor | 28 patients | No | No | Yes |
Jiang 2013 [146] | Classification | A | Non-medical | SVM-multi instance learning | 1 patient | No | No | No |
Jin 2016 [147] | Classification | PP | Non-medical | Improved random forest | 1 image, 1 patient | No | No | Yes |
Kashala 2020 [148] | Classification | A | Non-medical | FireNet module in SqueezeNet and obtained FCN as well-developed new particle swarm optimization called NPSO | LiTS 2017: 70 patients; 3D-IRCADb: cross-validation | No | No | Yes |
Khalili et al 2020 [149] | Classification | A | Non-medical | CNN | Validation: 2 patients; test: 2 × 10 patients | No | Yes | Yes |
Kumar 2013 [150] | Classification | A | Non-medical | Probabilistic neural network | 150 images | No | No | No |
Kutlu 2019 [151] | Classification | A | Non-medical | CNN with alexnet architecture, DWT (Discrete Wavelet Transform) and Long short-terms memory networks | 30% of data for test | No | No | No |
Yasaka et al 2018 [152] | Classification | A | Medical | CNN | 100 patients/image sets | Yes | Yes | No |
Xin et al 2020 [56] | Classification | A | Non-medical | CNN | 32 patients, 643 slices containing lesions | No | No | No |
Sreeja and Hariharan 2017 [153] | Classification | PP | Non-medical | SVM + Naive Bayes classifier | Not specified | No | No | No |
Shi et al 2020 [154] | Classification | A | Medical | CNN | One per lesion | No | No | No |
Selvathi et al 2013 [64] | Classification | PP | Non-medical | ELM + FCMC | Not specified | No | No | No |
Sayed 2016 [65] | Classification | PP | Non-medical | Fuzzy clustering + GWO (liver and liver lesion segmentation); SVM (liver disease classification: benign/malignant) | Not provided | No | No | No |
Romero et al 2019 [155] | Classification | PP | Non-medical | CNN (FCN × 2) | (1) 46 lesions for validation, 46 lesions for testing; (2) not specified | No | No | Yes |
Renukadevi and Karunakaran 2020 [156] | Classification | A | Non-medical | DBN + GOA | Not specified | Yes | No | Yes |
Rajathi 2019 [157] | Classification | A | Non-medical | WOA-SA + SVM + k-NN + RFC | 21 patients | No | No | No |
Raj 2016 [68] | Classification | PP | Non-medical | SVM | Not specified | No | No | No |
Ponnoprat et al 2020 [71] | Classification | A | Non-medical | U-Net for segmentation + CRF for post-processing + SVM for classification (w GHI kernel) | 17 patients, 2042 images | No | No | No |
Peng et al 2020 [158] | Classification | A | Medical | CNN (ResNet50) | ZHHAJU: 89; SYUCC: 138 patients | Yes | No | No |
Özyurt et al 2019 [159] | Classification | A | Non-medical | CNN | 34 | No | No | No |
Ouhmich et al 2019 [72] | Classification | A | Medical | U-Net | Not specified | No | No | No |
Nayak et al 2019 [74] | Classification | A | Medical | Segmentation: region-growing; classification: SVM | Not specified | No | No | Yes |
Mukherjee et al 2013 [75] | Classification | PP | Non-medical | SVM + PCA | Not specified | No | No | No |
Mofrad et al 2014 [78] | Classification | A | Medical | SVM (classification), k-NN (classification) | 1 patient | No | No | No |
Mala et al 2015 [160] | Classification | A | Non-medical | PNN, LVQ, BPN | 20 patients, ca. 20 images per patient | No | No | No |
Maaref et al 2020 [161] | Classification | A | Medical | 2D CNN (Inception-Net, modified) | CLASSIFICATION: 20 patients for validation, 41 for testing; PREDICTION: 12 patients for validation, 24 for testing | No | No | No |
Selvaraj 2013 [82] | Classification | PP | Non-medical | FCM (lesion segmentation) + BPSO (feature selection) + PNN (classification) | 15 images | No | No | No |
Li et al 2019 [162] | Classification | PP | Non-medical | BPN (+ PCA preprocessing) | 57 (tenfold cross-validation) | No | No | No |
Liang et al 2018 [163] | Classification | PP | Non-medical | CNN (ResNet w/ global and local pathways—for segmentation) + SVM (classification) | (1) Validation: 115, test: 96; (2) validation: 93, test: 110 | No | No | No |
Liang et al 2018 [163] | Classification | PP | Non-medical | CNN (ResNet w/ global and local pathways w/ bi-directional long short-term memory—for segmentation) + SVM (classification) | (1) Validation: 115, test: 96; (2) validation: 93, test: 110 | No | No | No |
Xin et al 2020 [56] | Classification | A | Medical | CNN | 643 slices for test | No | No | 0 |
Thuring et al 2020 [164] | Classification | A | Medical | Random forest and CNN | 70 patients | No | Yes | Yes |
Wang et al 2021 [165] | Classification | A | Medical | Nodule Net and HCCNet | 385 from same hospital, external test set with 556 patients | Yes | Yes | 0 |
Wang et al 2020 [166] | Classification | A | Non-medical | CNN (Siamese cross contrast neural network) | 67 patients | No | No | 0 |
Xu et al 2021 [167] | Classification | A | Medical | Random forest | tenfold cross-validation | No | No reference | 0 |
Zhang et al 2020 [168] | Classification | A | Medical | GLM | 57 | No | No | 0 |
Zhou et al 2021 [116] | Classification | A | Medical | CNN | 1/4 of lesion was used for test set | No | No | 0 |
Giannini et al 2020 [169] | Classification | A | Medical | Gaussian Naive Bayes classifier | 10 patients, 33 tumors/metastases | No | No | No |
Homayounieh et al 2020 [170] | Classification | A | Medical | Random forest | 103 patients w benign (60/103) or malignant (43/103) tumors | No | No | No |
Mao et al 2020 [171] | Classification | A | Medical | Gradient boosting (XGBoost) | 60 patients | No | No | No |
Mokrane et al 2020 [172] | Classification | A | Medical | Random forest | 36 patients | Yes | No reference | No |
Budai et al 2020 [173] | Miscellaneous | A | Medical | RF and SVM, K-means clustering | Independent validation dataset from > Sliver07(20 dataset), > MICCAI 2017 (LiTS) 131 scans | No | No reference | Yes |
Choi et al 2018 [33] | Miscellaneous | A | Medical | CNN | 150 images | Yes | Yes | No |
Huo et al 2019 [174] | Miscellaneous | A | Medical | DCNN and morphological operation for attenuation and SS-Net (a DCNN model) | Not specified | Yes | Yes | Yes |
Kayaalti et al 2014 [175] | Miscellaneous | A | Non-medical | SVM and K-nearest neighbors for classifying the images | No | No | No | |
Yasaka et al 2018 [176] | Miscellaneous | A | Medical | CNN | 100 portal phase images from 100 patients | No | Yes | No |
Son et al 2020 [177] | Miscellaneous | A | Medical | U-net | Not specified | No | Yes | No |
Yin et al 2021 [178] | Miscellaneous | A | Medical | CNN | fivefold cross-validation | No | No | 0 |
Ahmadi et al 2016 [179] | Miscellaneous | A | Medical | FCM and GA | Test dataset 1: 150 patients, test dataset 2: 50 patients | No | No | No |
Ben-Cohen et al 2018 [180] | Miscellaneous | PP | Non-medical | U-net base—using unlabeled data features in supervised network | Test set 1: 421 patients. Test set 2: 298 (other institutions). Test set 3: 172 patients (from tertiary referral hospitals | No | No | No |
Bevilacqua et al 2017 [27] | Miscellaneous | PP | Non-medical | ANN classifier by using mono-objective genetic algorithm (GA) | Not mentioned | Yes | No reference | No |
Conze et al 2017 [181] | Miscellaneous | A | Medical | Scale adaptive super voxel-based random forests | Not specified | No | No reference | No |
Fu et al 2019 [6] | Miscellaneous | A | Non-medical | U net-with multi stream feature fusion and multi scale dilated convolution, author called it M-Net | Not specified | No | No | No |
Gensure et al 2012 [182] | Miscellaneous | A | Medical | SVM | Not provided | No | No | No |
Huang et al 2018 [183] | Miscellaneous | A | Medical | 3d U-Net | Not specified | Yes | Yes | Yes |
Kumar et al 2016 [50] | Miscellaneous | A | Medical | SVM, weighted nearest neighbor | 50 CT images were used from ImageCLEF 2014 with tenfold cross-validation | No | No | Yes |
Zhang et al 2018 [52] | Miscellaneous | A | Non-medical | Fuzzy connectedness (fuzzy logic) | (1) VascuSynth: not eligible; (2) 3D-IRCADb: not provided; (3) Sliver07: 10 patients | Yes | No | Yes |
Zeng et al 2016 [184] | Miscellaneous | A | Medical | ELM | 100,000 images in total (training + testing data) | No | No | No |
Yu et al 2019 [185] | Miscellaneous | PP | Non-medical | CNN | 6 cases (+ 3 for validation); slices per case range: 135–500 | No | No | No |
Yang et al 2012 [186] | Miscellaneous | A | Medical | k-means | Not specified | No | No | No |
Xin et al 2020 [56] | Miscellaneous | A | Non-medical | CNN | 32 patients, 643 slices containing lesions | No | No | No |
Wang et al 2018 [59] | Miscellaneous | PP | Non-medical | BoVW (K-CP with multilinear OMP, K-nearest neighbor) | Leave-on-out cross-validation is used for testing | No | No | No |
Taghavi et al 2021 [9] | Miscellaneous | A | Medical | Random forest | 21 patients | No | Yes | No |
Ponnoprat et al 2020 [71] | Miscellaneous | A | Non-medical | U-Net for segmentation + CRF for post-processing + SVM for classification (w GHI kernel) | 17 patients, 2042 images | No | No | No |
Maaref et al 2020 [161] | Miscellaneous | A | Medical | 2D CNN (Inception-Net, modified) | CLASSIFICATION: 20 patients for validation, 41 for testing; PREDICTION: 12 patients for validation, 24 for testing | No | No | No |
Wang et al 2017 [187] | Miscellaneous | A | Non-medical | BoVW (sparse codebook-based feature representation) | (leave-one-out cross validation) | No | No | No |
Li et al 2020 [188] | Miscellaneous | A | Medical | ResNet | 69 patients, 3 images per patient (fivefold cross-validation) | No | No | No |
Lee et al 2020 [8] | Miscellaneous | A | Non-medical | CNN + RFC and CNN + LRC | 606 patients | No | No | No |
Sun et al 2020 [189] | Miscellaneous | PP | Non-medical | SVM | 34 labeled CT | No | No | 0 |
Thuring et al 2020 [164] | Miscellaneous | A | Medical | Random Forrest & CNN | 70 patients | No | Yes | Yes |
Wang et al 2020 [166] | Miscellaneous | A | Non-medical | CNN (residual CNN) | 70slices (17 patients) | No | No | 0 |
Xu et al 2020 [190] | Miscellaneous | PP | Non-medical | CNN (Deep neural network) | 20 from 3dIRCADb | Yes | No reference | Yes |
Yang et al 2021 [191] | Miscellaneous | A | Non-medical | CNN (v-net) | 8 CT | No | No | Yes |
Yoshinobu et al 2020 [192] | Miscellaneous | PP | Non-medical | CNN (Deep CNN) | 32 cases | No | No | 0 |
Zhang et al 2020 [124] | Miscellaneous | A | Medical | CNN (DenseNet) | From multicenter data from 3 hospitals | Yes | Yes | Yes |
Gu et al 2020 [193] | Miscellaneous | PP | Non-medical | CNN + ResNet | 1 patient | No | No | No |
Kobe et al 2021 [194] | Miscellaneous | A | Medical | ANN | 21 metastases/lesions | No | No reference | No |
Li et al 2022 [195] | Miscellaneous | A | Medical | CNN (DenseNet) | 244 patients | No | Yes | Yes |
Liver segmentation
Segmentation | refers to a pixel-wise classification of images throughout this review. This is the standard meaning of “segmentation” of images in data science and engineering. It is not to be confused with anatomical segmentation like the Coineaud segmentation of liver lobes, commonly used for clinical segmentation of the liver according to the portal blood supply (19) |
DICE | describes the percentage of overlap between the predicted and the observed/”correct” labeled area in an image (often labeled by a human radiologist), where 1.0/100% means a perfect overlap between predicted and correct segmentation |
Accuracy | related to image segmentation in engineering is a measure describing how many pixels are correctly classified—1.0/100% being perfect. However, accuracy can be misleading in cases where a class is in very few pixels; for instance, a small tumor could be only in 2% of the image—and a model predicting that there are 0% tumors would still have an accuracy of 98%. Therefore, if only accuracy is reported for performance, a measure of class balance might be relevant to the readers' understanding |
Precision and Recall | Precision is the number of relevant observations by a model divided by the total number of observations made by the model. For instance, if a model marks 100 pixels as tumor tissue and 40 are tumor tissue, the precision is 40%/0.4. Precision is the same as positive predictive value (PPV). Recall is the number of relevant observations divided by the total number of actual cases, e.g., if an image contains 100 pixels with actual tumor tissue, and the model observes 80 of them, the model has a recall of 80%/0.8. In binary classification cases, recall is the same as sensitivity, hit rate, and true positive rate |
Volume Overlap Error (VOE) | gives a measure of the difference between actual area and predicted area. It functions as a combined score of both false positives and negatives \({\varvec{V}}{\varvec{O}}{\varvec{E}}\left({{\varvec{U}}}_{1},{{\varvec{U}}}_{2}\right)=100\times \boldsymbol{ }(1-\boldsymbol{ }\frac{{{\varvec{U}}}_{1}\cap \boldsymbol{ }{{\varvec{U}}}_{2}}{{{\varvec{U}}}_{1}\cup \boldsymbol{ }{{\varvec{U}}}_{2}})\) where U1 and U2 are true and predicted values, respectively. Optimal scores are as low as possible, 0 being the perfect score (20) |
IoU / Jaccard Index | The intersection over union (IoU), is a measure that quantifies the percentage of overlap between prediction and observed/true output, much like the DICE coefficient. IoU measures the overlapping pixels between true and predicted segmentation and divides it by the total number of pixels either of them has marked as a pixel of interest. A perfect score would be 100%/1.0. This measure is also referred to as the Jaccard Index |
Ground truth | refers to the label for anatomical structures in CT images given by a clinician or radiologist. What kind of expert and level of experience is often specified in each specific study |
CNN | refers to Convolutional Neural Network – a deep learning model based on vector calculations used in image recognition and processing pixel data |
Characteristics of studies | Liver segmentation | Lesion segmentation | Lesion detection | Classification of liver or lesions | Miscellaneous |
---|---|---|---|---|---|
Journal article | |||||
Proceeding papers | |||||
ML to human expert | |||||
Using public datasets | |||||
Reporting of standard error | |||||
Reporting of DICE score | |||||
Reporting of accuracy | |||||
Reporting of AUC | |||||
Reporting of precision | |||||
Reporting of VOE | Not available | Not available | 1 study [27] | ||
External validation |