Table 1.
List of included articles on RFs derived of MRI. In the second column # are the number of patients enrolled retrospectively (R) or prospectively (P). In the fourth column the volume of interest (VOI) is presented accompanied by the type of segmentation in brackets M = manual, SA = semiautomatic and A = fully automatic. The last column contains information on validation. “e” stands for external validation and “I” for internal. The number stands for the number of cohorts used. 2 means one for development and one for testing.
| Prostate cancer detection | ||||||
|---|---|---|---|---|---|---|
| Study | # | Imaging Modality | VOI (Segmentation) | Endpoint(s) | Results | Validation |
| Cameron et al. 67 | 5 (R) | T2w, ADC | PCa (A +M) | PCa detection | RF model outperformed conventional mpMRI feature models. | i (LOO) |
| Cameron et al. 66 | 13 (R) | T2w, DWI, DCE | PCa (A) | Classifiers for PCa detection | RF model outperformed conventional mpMRI feature models. | i (CV, LOO) |
| Viswanath et al. 68 | 85 (R) | T2w | PCa, PZ, central gland (M) | Classifier for voxel-wise PCa detection | Boosted Decision Tree classifier has the highest ROC-AUC for detecting PCa., Boosted Quadratic-Discriminant Analysis is the most accurate and robust in detection of PCa extent across three sites. The ground truth was established by whole mount histology. | e (CV with external centers) |
| Khalvati et al. 69 | 20 (R ) | T2w, DWI, CDI, CHB-DWI | PCa, prostate (M) | Classifier for PCa detection | Support vector machine classifier improved PCa auto-detection. | i (LOO) |
| Xu et al. 70 | 331 (R) | T2w, DWI, ADC | PCa | Benign vs. malignant lesions | BpMRI improved discrimination between benign and malignant lesions. | i (2) |
| Bonekamp et al. 71 | 316 (R) | T2w, DWI and ADC | PCa, PZ, prostate (M) | PCa ISUP ≥2 | Quantitative ADC measurement improves differentiation of benign vs malignant lesions, ML comparable, performance of zone-specific models was lower. | i (2+CV) |
| Sidhu et al. 72 | 76 (R) | T1w, T2w, DWI, ADC | PCa in TZ (M) | PCa detection in TZ | TZ derived RF can discriminate TZ-PCa. | i (LOO) |
| Ginsburg et al. 73 | 80 (R) | T2w, DWI, DCE | PCa, TZ, prostate (M) | PCa detection in TZ and PZ | TZ-specific classifier significantly improves accuracy of PZ-PCa detection. | e (3 institutions) |
| Parra et al.114 | 52 (R) | DCE | Habitat = biopsy +15 mm (M) | PCa detection (significant) | Habitats from DCE predict clinically significant PCa well. | i (LOO) |
| Khalvati et al. 74 | 30 (R) | T2w, ADC, CHB-DW, CDI | PCa (A) | Framework for PCa detection | Proposed framework (MPCaD can be utilized to detect and localize PCa. | i (LOO) |
| Wang et al. 75 | 54 (R) | T2w, CHB-DW | PCa, histological-radiological correlation (M) | Classifier for PCa detection (significant) | SVM classifier improves performance of PI-RADS v2 for clinically relevant PCa. | i (LOO) |
| Gholizadeh et al.76 | 16 (P) | T2WI, DWI, DTI | PCa in PZ(M) | Differentiation pf PCa and non-PCa | Voxel‐based supervised machine learning models generated a binary classification of cancer probability maps. | i (2+LOO) |
| Hu et al. 77 |
136 (P) | DWI, ADC | PCa (M) | PCa detection | A mixed model based on the clinically independent risk factors and mp-MRI radiomics score showed the best performance. | i (2) |
| Woźnicki et al.78 | 191 (R) | T2w, ADC | PCa + Prostate (M) | PCa detection and clinical significance | An ensemble machine learning model combining radiomics, PI-RADS, prostate specific antigen density and digital rectal examination resulted in a good predictive performance. |
i (2+CV) |
| Qi et al. 79 |
199 (R) | T2w, ADC, DCE | PCa (M) | PCa prediction on patients with PSA level of 4-10ng/ml | The combined model incorporating all sequences, age, PSA density and the PI‐RADS v2 score yielded good performance for prediction of PCa. | i (2) |
| Dulhanty et al. 115 | 101 (R) | ADC, CHB-DWI | Prostate zones (M) | PCa detection based on 10 anatomical zones | Zone-level radiomic sequences distinguish between positive and negative zones. | i (CV) |
| Bleker et al. 80 |
206 (P) | T2w, DWI, ADC, DCE | PCa (SA) | csPCa in PZ | Addition of DCE-RFs does not improve performance of T2w- and DWI-RF based models. Multivariate RF selection with extreme gradient boosting outperformed univariate selection. | i (2) |
| Wu et al.81 | 90 (R) | T2w, ADC | PCa in TZ (M) | Differentiation of PCa inTZ | Proposed models using quantitative ADC, shape and texture features, show good performance for TZ PCa detection and remained accurate when comparing TZ PCa with stromal BPH and in smaller lesions. | i (CV) |
| Kwon et al. 82 | 344 (R | T2w, DCE, DWI, proton density-weighted | Prostate, PZ, PCa (M) | Detection of csPCA Classification methods |
Random forest classification showed the highest AUC. | i (2) |
| Gleason score | ||||||
| Hectors et al.83 | 64 (R) | T2w, ADC, diffusion kurtosis imaging maps | PCa (unknown) | Aggressiveness (GS, Gene expression, Decipher) | 14 RF with significant correlation to GS, 40 DWI features with significant correlation to Gene expression, ML models with excellent performance to predict Decipher score ≥ 6. | i (CV) |
| Chaddad et al. 85 | 99 (R) | T2w, ADC | PCa (M) | GS grouping (6/3+4/4+3) | Joint Intensity Matrix-derived RF (n=5) are independent predictors of GS. | i (2) |
| Chaddad et al. 84 | 99 (R) | T2w, ADC | PCa (A) | GS grouping (6/3+4/4+3) | T2w and ADC derived RF can predict GS. | i (CV) |
| Sun et al. 119 | 30 (R) | T2w | PCa on histology (M) | GS, Risk groups | ADC, GLCM and GLRLM discriminate between high grade and low grade PCa. The combination further improved AUC. | i (CV) |
| Jensen et al. 86 | 112 (R) | T2w, DWI | PCa (M) | GS, risk group | Zonal-specific DWI and T2w derived RF differentiate between PCa lesions of all GS. | i (LOO + CV) |
| Chen et al. 87 | 381 (R) | ADC, T2w | PCa, prostate(M) | PCa/non-PCa, high grade GS /low grade GS 6 compared to PI-RADSv2 | T2w and ADC RF show high efficacy in distinguishing PCa vs non-PCa and high-grade vs low-grade PCa. | i (2) |
| Toivonen et al. 88 | 62 (R) | T2w, DWI, T2-mapping | PCa | GS | T2w and DWI derived RF show good classification performance for GS of PCa. | i (LPOCV + CV) |
| Zhang et al. 89 | 166 (R) | T2w, ADC, DCE | PCa (M) | PCa upgrading | T2w, ADC and DCE derived RF can predict GS upgrading from biopsy to radical prostatectomy. | i (2) |
| Min et al. 90 | 280 (R) | T2w, DWI, ADC | PCa (M) | PCa detection (significant) | MpMRI derived RF discriminate between GS 3+4 or lower. | I (CV) |
| Li et al. 91 | 63 (R) | T2w, ADC, DCE | PCa (M) | GS in CG PCa | Support vector machine classification achieves accurate GS classification of PCa in central gland. | i (CV) |
| Rozenberg et al. 92 | 54 (R) | ADC | PCa (M) | Prediction of GS upgrading and Differentiation of GS 3+4 and 4+3 | ADC derived texture features are not predictive of GS upgrading after radical prostatectomy. | i (CV) |
| McGarry et al. 120 | 48 (P) | T2w, ADC, DCE | PCa on histology (M) | Gleason probability maps | RF based mapping successfully stratifies high- and low-risk PCa. | i (2) |
| Penzias et al. 121 | 36 (R) | T2w | PCa on histology (M) | GS, risk group, correlation with QH | RF and quantitative histomorphometry features correlated with these RF are predictive for of GS. | i (2) |
| Fehr et al. 93 | 217 (R) | T2w, ADC | PCa (M) | GS risk group differentiation | Automatic classifiers achieve accurate classification of GS. | i (CV) |
| Hou et al. 94 | 263 (R) | T2w, DWI, ADC | PCa, (M) | Clincially significant PCa (GS≥7) in PIRADS 3 lesions | Radiomics ML model of all sequences has potential to predict csPCa in PIRADS 3 lesions to guide biopsy. | i (CV) |
| Li et al. 95 | 381 (R) | T2w, ADC | PCa in TZ and PZ (M) | Clincically significant PCa | Radiomics model can predict csPCa with high accuracy (AUC ≥-98). | i (2) |
| Gong et al.116 | 489 (R) | T2w, DWI, ADC | Prostate | Identification of high grade PCa (>GS7) | DWI RF-model and combination of T2w and DWI achieved high accuracy in prediction of GS >7. | i (2, CV) |
| Algohary et al. 96 | 231 (R) | T2w, ADC | PCa lesion, peritumoral area (M) | Differentiation of PCa Risk Groups according to D'Amico | Combination of peritumoral and intratumoral RFs improved the risk stratification results by 3-6% compared to intra-tumoral features alone. | e (2) |
| Gugliandolo et al. 97 | 65 (R) | T2w | Prostate excluding urethra and dominant intraprostatic lesions (M) | Prediction of GS, PIRADS v2 Score and Risk Group | Radiomic signature consisting of the combination of 3D GLCM and intensity domain category features were able to discriminate between low- and intermediate-grade malignancy. | i (CV, LOO) |
| Zhang et al. 98 | 159 (R) | T2w, DWI, ADC | PCa (M) | Discrimintation of csPCa and clincially insignificant PCa | A radiomic signature of 10 features, was significantly associated with csPCa. A nomogram of this signature and ADC values showed even better AUCs. | e (2, CV) |
| Algohary et al. 99 | 56 (R) | T2w, ADC | PCa (M) | Prediction of csPCa in active surveillance patients | 7 T2w-based and 3 ADC-based RF exhibited statistically significant differences between malignant and normal regions in the training groups. The 3 constructed ML models yielded good accuracy | i (CV) |
| Abraham et al. 100 | 162 (R) | T2w, ADC, high B-Value Diffusion-Weighted (BVAL) | PCa (A) | Classification of Grade Groups | The novel method using texture features and stacked sparse autoencoder was able to classify PCa grade groups moderately. | i (2, CV) |
| Extracapsular extension | ||||||
| Ma et al. 120 | 119 (R) | T2w | PCa (M) | ECE of PCa | T2w derived RF predict side specific ECE. | i (2) |
| Ma et al. 90 | 210 (R) | T2w | PCa (M) | ECE prior to RP | T2w derived RF outperformed radiologist in predicting ECE. | i (2) |
| Stanzione et al. 103 | 39(R) | T2w, AdC | PCa index Lesions (M) | Classifier for ECE prediction | Bayesian Network was the best classifier for ECE prediction. | i (CV) |
| Losnegard et al. 104 | 228 (R) | T2w, ADC, DCE | Prostate, PCa (M+A) | ECE Prediction in high and unfav. Intermediate risk PCa | 12 RF extracted from manual segmentation combined with a Random Forest classifier can predict ECE with an AUC of 0.74. Features from T2W and ADC showed a good performance. A combined model performed even better. |
i (CV) |
| Xu et al. 105 | 95 (R) | T2w, DWI, ADC, DCE | PCa (M) | ECE | 8 RF were used to build a radiomics model with an AUC of 0.92. A radiomics nomogram with clinical features yielded similar results. | i (2) |
| Bone metastasis | ||||||
| Wang et al. 106 | 176 (R) | T2w, DCE T1w | PCa (M) | Bone metastasis prediction | T2w and DCE derived RF were predictors for BM. | i (2) |
| Zhang et al. 107 | 116 (R) | T2, DWI, DCE | PCa (M) | Prediction of bone metastasis in newly diagnosed PCa | The radiomics nomogram based on 11 RFs and clinical risk factors, showed good performance to promote individualized prediction of bone metastasis. | i (2) |
| Biochemical recurrence | ||||||
| Bourbonne et al. 109 | 107 (R) | T2w, ADC | PCa (SA) | Prediction of BCR and biochemical relapse free survival after RP in high risk PCa | One ADC derived RF (SZEGLSZM) was predictive for BCR and bRFS (AUC 0.76). | i (2) |
| Bourbonne et al. 108 | 195 (R) | ADC | PCa (SA) | BCR | External validation of the identified ADC derived RF (SZEGLSZM) for BCR and bRFS prediction after RP. | e (2) |
| Shiradka et al. 110 | 120 (R) | T2w and ADc | PCa, prostate (M) | BCR after RP or RT | BpMRI RF-trained machine learning classifier can be predictive of BCR. | e (2) |
| Zhong et al.117 | 91 (R) | T1w, T2w, DWI | Prostate (M) | BRC of localized PCa after RT and neoadjuvant endocrine therapy. | MRI derived RFs can predict BCR after RT with good performance. | i (2, CV) |
| Treatment response | ||||||
| Abdollahi et al. 111 | 33 (P) | T2w, ADC, pre- and post IMRT | PCa (M) | Therapy response (RT), GS, T-stage | T2w and ADC derived RF and ML correlate with IMRT response. | i (CV) |
| Toxicity | ||||||
| Abdollahi et al. 122 | 33 (P) | T2w, ADC | Rectal wall (M) | Rectal toxicity | Pre-IMRT MRI RF predict rectal toxicity. | i (CV) |
| Segmentation | ||||||
| Sunoqrot et al. 118 | 635 (R) | T2w | Prostate gland (M) | Quality System for automated prostate segmentation | Proposal of a quality check for automated segmentation of the prostate in T2W MR image. | e (2, CV) |
| Lay et al. 112 |
224 (R) | T2w, ADC, DWI | PCa (M) Prostate and TZ (A) | PCa segmentation | Random forest sampling strategy and instance-level weighting improve PCa detection performance compared to support vector machine. | i (2, CV) |
| Giannini et al. 113 |
58 (R) | T2w, ADC | PCa (M) | PCa segmentation | Proposed method with GLCM texture features computed on ADC and T2w images reduced the number of false positives and increased the precision of PCa detection. | i (CV) |
Abbreviations: ADC=Apparent diffusion coefficient, BCR=biochemical recurrence, bpMRI=biparametric magnetic resonance imaging, bRFS=biochemical recurrence free survival, CDI=current density imaging, csPCa= clinically significant prostate cancer, CV=cross validation, DCE=dynamic contrast enhanced, DTI= diffusion. tensor imaging, DWI=diffusion weighted imaging, GLCM= gray level co-occurrence matrix, GLRLM=grey-level run length matrix, GS=Gleason score, IMRT=intensity modulated radiotherapy, LOO=leave one out, LPOCV=leave-pair-out cross-validation, M=manual confirmation, ML=machine learning, mpMRI=multiparametric magnetic resonance imaging, PCa=Prostate cancer; PZ=peripheral zone, RF=radiomic feature, ROC-AUC=are under the receiver operating characteristics curve, RP=radical prostatectomy, T1w= T1-weighted imaging, T2w=T2-weighted imaging, TZ= transitional zone.