Table 4.
Predictive modelling characteristics of studies using traditional machine learning-based semi-automated AI methods
| Study | No. of patients | Training set | Validation set | Test set | Algorithm | MRI input | IR | IS | Discriminative features | No. of features used for training | Outcome | Zone | Analysis | Evaluation strategy |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Bonekamp [25] | 316 | 183 | NR | 133 | RF | T2WI, ADC, b = 1500 | No | Manual | First-order, volume, shape, texture | NR | csPCa vs iPCa or benign lesions | WP or PZ or TZ | Per lesion and per patient | Internal hold-out |
| Min [26] | 280 | 187 | NR | 93 | LR | T2WI, ADC, b = 1500 | No | Manual | Intensity, shape, texture, wavelet | 9 | csPCa vs iPCa | WP | Per lesion | Internal hold-out |
| Kwon [27] | 344 | 204 | tenfold CV | 140 | CART, RF, LASSO | T2WI, DWI, ADC, DCE | Rigid | No | Intensity | 54 | csPCa vs iPCa or benign lesions | PZ or TZ | Per lesion | Internal hold-out |
| Castillo [28] | 107 | 80% | 20% of training (100 random repeats) | 20% | LR, SVM, RF, NB, LQDA | T2WI, DWI, ADC | HPa | Manual | Shape, local binary patterns, GLCM | NR | csPCa vs iPCa | WP | Per lesion, Per patient | Mixed hold-out |
| Bleker [29] | 206 | 130 | NR | 76 | RF, XGBoost | T2WI, b = 50, b = 400, b = 800, b = 1400, ADC, Ktrans | No | Manual | Intensity, texture | NR | csPCa vs iPCa or benign lesions | PZ | Per lesion | Internal hold-out |
| Li [30] | 381 | 229 | NR | 152 | LR | T2WI, ADC | No | Manual | Intensity, age, PSA, PSAd | 15 | csPCa vs iPCa or benign lesions | WP | Per lesion | Internal hold-out |
| Woźnicki [31] | 191 | 151 | fivefold CV | 40 | LR, SVM, RF, XGBoost, CNN | T2WI, ADC | No | Manual | Intensity, shape, PI-RADS, PSAd, DRE | 15 | csPCa vs iPCa or benign lesions | WP | Per patient | Internal hold-out |
| Bevilacqua [32] | 76 | 48 | threefold CV | 28 | SVM | ADC, b = 2000 | No | Manual | Intensity | 10 | csPCa vs iPCa | WP | Per lesion | Internal hold-out |
| Toivonen [33] | 62 | 62 | LPOCV | N/A | LR | T2WI, ADC, Ktrans, T2 map | No | Manual | Intensity, Sobel, texture | NR | csPCa vs iPCa | WP | Per lesion | LPOCV |
| Antonelli [34] | 164 | 134 | NR | 30 |
PZ: LinR TZ: NB |
ADC, DCE | Rigid | Manual | Texture, PSAd | NR | csPCa vs iPCa | PZ or TZ | Per lesion | fivefold CV |
| Yoo [35] | 427 | 271 | 48 | 108 | CNN, RF | ADC, DWI | No | No | First-order statistics of deep features | 90 | csPCa vs iPCa or benign lesions | WP | Per slice, Per patient | tenfold CV |
| Hiremath [36] | 592 | 368 | threefold CV | 224 | AlexNet or DenseNet and Nomogram | T2WI, ADC | Rigid, affine | Manual | Deep learning imaging predictor, PI-RADS, PSA, gland volume, tumour volume | NR | csPCa vs iPCa or benign lesions | WP | Per patient | External hold-out |
ADC, apparent diffusion coefficient; CART, classification and regression trees; CNN, convolutional neural networks; GLCM, grey level co-occurrence matrix; HP, histopathology; IR, image registration; IS, image segmentation; LASSO, least absolute shrinkage and selection operator; LinR, linear regression; LQDA, linear and quadratic discriminant analysis; LR, logistic regression; NB, naïve Bayes; PI-RADS, prostate imaging-reporting and data system; PSA, prostate-specific antigen; PSAd, prostate-specific antigen density; RF, random forests; SVM, support-vector machines
aHistopathology images registered with T2-weighted images using specialised software