Table 2.
Main characteristics of the machine learning process of the selected studies.
| Reference, Year | Sequences | Segmentation | Feature Extraction | Image Preprocessing | Data Imbalance techniques, Data Augmentation | Feature Selection | Train/Test (%) b | Algorithm |
|---|---|---|---|---|---|---|---|---|
| Dominguez et al. 2023 [24] | T2, ADC | Lesion | Shape, FO, HTF | Not performed | NR | RFE | 80 (CV)/20 | LR |
| Prata et al. 2023 [25] | T2, ADC | Lesion | FO, HTF, BLP | NR | NR | Wrapper (RF) | CV | RF |
| Jin et al. 2023 [26] | T2, ADC, DWI (b2000) | Lesion | FO, HTF, wavelet features | IN, grey-level quantization, resampling, IR | SMOTE, NR | ANOVA | 70/30 | SVM |
| Hamm et al. 2023 [27] | T2, ADC, DWI (high-b value) | Lesion, prostate, PZ, TZ | Deep radiomics | IN, resampling, lesion cropping | NR, Yes | NA | 80 (CV)/20 | Visual Geometry Group Net-based CNN |
| Hong et al. 2023 [28] | ADC | Lesion, prostate | Deep radiomics | IN, resizing, prostate cropping, cut-off filtering | Image allocation, NR | NA | 80/20 | DenseNet 201 |
| Jing et al. 2022 [29] | T2, DWI (b1500) | Lesion, prostate | Shape, FO, HTF, higher-order features | IN, Resampling | NR | Variance threshold algorithm, Select K-best, LASSO | 70/30 | LR |
| Zhu et al. 2022 [30] | T2, ADC | Lesion | Deep radiomics | IN, resampling, prostate cropping, IR | NR, Yes | NA | 60/40 | Res-UNet |
| Jiang et al. 2022 [31] | T2, DWI (b1500), ADC | Lesion, prostate | Deep radiomics | IN, resampling, prostate cropping, IR | NR, Yes | NA | 66.6/33.3 | Attention-Gated TrumpetNet |
| Liu et al. 2021 [32] | T2, ADC | Lesion | Deep radiomics | IN, lesion cropping, IR | NR | NA | 70/30 | 3D GLCM extractor + CNN |
| Lim et al. 2021 [33] | T2, ADC | Lesion | Shape, FO, HTF | NR | NR | Mann–Whitney U-test | CV | XGBoost |
| Hectors et al. 2021 [34] | T2 | Lesion | Shape, FO, HTF | IN, grey-level quantization, resampling | SMOTE, NR | RF | 80 (CV)/20 | RF, LR |
| Castillo et al. 2021 [35] | T2, DWI (highest-b value), ADC | Lesion | Shape, FO, HTF, higher-order features | Resampling | WORC Workflow a | WORC Workflow a | 80 (CV)/20 | WORC Workflow a |
| Li et al. 2020 [36] | T2, ADC | Lesion | FO, HTF | IN, grey-level quantization, resampling | NR | mRMR, LASSO | 60/40 | LR |
| Zhong et al. 2019 [37] | T2, ADC | Lesion | Deep radiomics | IN, resizing, lesion cropping | Not necessary, Yes | NA | 80/20 | ResNet with TL |
BLP = binary local pattern, CNN = convolutional neural network, CV = cross-validation, FO = first order, GLCM = gray-level co-occurrence matrix, HTF = handcrafted texture features, IN = image normalization, IR = image registration, LASSO = least absolute shrinkage and selection operator, LR = logistic regression, mRMR = minimum redundancy maximum relevance, NR = not reported, NA = not applicable, PZ = peripheral zone, RFE = recursive feature elimination, RF = random forest, SMOTE = synthetic minority oversampling technique, SVM = support vector machine, TZ = transitional zone. a Uses a radiomics workflow called Workflow for Optimal Radiomics Classification (WORC), which includes different workflow processes (see reference for further details). b Presented as % of the data selected for the training and test partitions. CV stands for cross-validation performed in the training set.