Table 1.
Clinical benefits of radiomic and radiogenomics in CRC liver metastatic patients.
| Study | Design | Imaging Modalities | Sample Size | Study Cohorts and Validation | Tools for Radiomics Calculations | Statistical Model Construction |
|---|---|---|---|---|---|---|
| Early diagnosis of colorectal cancer metastasis | ||||||
| Becker et al., 2018 [5] | Preclinical | MRI | 8 male mice | One cohort | MATLAB routine | Linear regression model, Pearson correlation test and hierarchical cluster analysis |
| Taghavi et al., 2021 [6] | Retrospective | CT | 91 CRC without LM at diagnosis | Two cohorts. Patients with metastases in follow-up of ≥24 months (n = 67); and patients who developed metachronous liver metastases <24 months (n = 24). No validation |
Philips Intellispace Portal software and PyRadiomics | Kruskal–Wallis test, inter-correlated features and Bayesian-optimized random forest was used for prediction models. |
| Rao et al., 2014 [7] | Retrospective | CT | 29 CRC patients | Three cohorts. Patients without LM (n = 15), with synchronous LM (n = 10) and metachronous LM within 18 months following primary staging (n = 4). No validation |
MATLAB routine | Student’s t test or Mann–Whitney U test. ROC analyses to determine the potential diagnostic performance of the respective texture parameters for diagnosing the presence of metastatic disease. |
| Liang et al., 2019 [8] | Retrospective | MRI | 108 rectal cancer patients | Two cohorts. 54 patients with LM and 54 without LM. The results of the one-round cross-validation were stabilized and representative. |
Python in Anaconda3 platform with Scikit-learn and Matplotlib packages. | Models were evaluated with indicators of accuracy, sensitivity, specificity and AUC, and compared by DeLong test. |
| Oyama et al., 2019 [9] | Retrospective | MRI | 150 liver tumors. 50 HCC, 50 LM and 50 HHs in 37, 23 and 33 patients | One cohort. | MATLAB Image Processing Toolbox, Signal Processing Toolbox, Statistics and Machine Learning Toolbox, and Wavelet Toolbox | Two machine learning models: a logistic classifier model with an elastic net penalty and extreme gradient boosting (XGBoost) |
| Li et al., 2017 [10] | Retrospective | MRI | 162 patients | Three cohorts. HHs (n = 55), LM (n = 67) and HCC (n = 40). The test datasets validated the reliability of the models |
R software (R Core Team, Vienna, Austria) and MATLAB R2013b (Mathworks, Natick, MA, USA) | Kruskal–Walls test, ROC curve and AUC analysis to differentiate three subtypes. K-nearest neighbor classifier model, back-propagation artificial neural network classifier model, support vector machine and logistic regression were used for improving accuracy for classifier. |
| Jansen et al., 2019 [11] | Retrospective | MRI | 95 patients with 125 benign lesions and 88 malignant lesions | Two cohorts, benign and malignant lesions. 40 adenomas, 29 cysts and 56 HHs; and 30 HCC and 58 LM. Optimization process using cross-validation. |
- | ANOVA F-score was selected and fed into an extremely randomized trees classifier and ROC curve analysis. |
| Gatos et al., 2017 [12] | Retrospective | MRI | 71 FLLs. 30 benign lesions and 41 malignant lesions | Three cohorts. 30 benign lesions, 19 HCC and 22 LM. No validation |
- | Probabilistic Neural Network (PNN) model evaluation was performed using the leave-one-out (LOO) method and receiver operating characteristic (ROC) curve analysis. Multilinear regression analysis. |
| Response assessment and treatment decision tool | ||||||
| Taghavi et al., 2021 [13] | Retrospective | CT | 90 CRC patients with 140 LM treated by ablation | Two cohorts. Training (n = 63 patients/n = 94 lesions) and validation (n = 27 patients/n = 46 lesions) cohort. Each patient was considered as one group in the fivefold cross-validation to ensure that all lesions for each patient were in the training/test set of a fold |
3D slicer and 3D using the Pyradiomics package in Python (3.7) | Three models: each model was based on a Cox’s proportional hazards model. |
| Staal et al., 2021 [14] | Retrospective | CT | 82 CRC patients with 127 LM treated by ablation | One cohort. Internal validation. |
- | Kruskal–Wallis test was applied to evaluate whether the selected radiomics features were influenced by differences between scanners. Combined model yielded a c-statistic. Multivariable Cox regression |
| Reimer et al., 2018 [15] | Retrospective | MRI | 37 CRLM patients treated by TARE | One cohort. | Mint Lesion ™ 3.0 (Mint Medical GmbH, Dossenheim, Germany) | Mann–Whitney U test. AUC and sensitivity and specificity were calculated. |
| Shuer et al., 2019 [16] | Retrospective | CT and MRI | 102 CRLM treated by resection | One cohort. | Pyradiomics plugin to 3D Slicer | Cox regression coefficients |
| Ahm et al., 2016 [17] | Retrospective | CT including quadruple-phase (n = 27), triple-phase (n = 141), double-phase (n = 11) and single-phase CT (n = 54) | 145 patients | Two cohorts. Validation cohorts (n = 90) and derivation cohorts (n = 145). | In-house software program (Medical Imaging Solution for Segmentation and Texture Analysis). | Student t, Mann–Whitney U test, ×2 or Fisher exact test. Multivariate logistic regression analysis. |
| Giannini et al., 2020 [18] | Included in HERACLES trial | CT | 38 patients | Two cohorts. Training cohort 28 patients (108 lesions), validation cohort 10 patients (33 lesions). | Mipav software. In-house framework based on C++ and libraries | Genetic algorithms, algorithms belonging to the computational intelligence field. |
| Beckers et al., 2018 [19] | Retrospective | CT | 70 CRLM patients | Two cohorts. 60 patients with chemotherapy and 10 patients without chemotherapy. No validation. |
2D Texture analysis was performed with in-house software written in Python (MANGO; Multi-image Analysis GUI, Research Imaging Institute). | Shapiro–Wilk test was used to test for normality. Independent sample t tests. Multivariable Cox proportional hazards models |
| Andersen et al., 2019 [20] | Exploratory study | CT | 27 CRLM patients treated by regorafenib | One cohort | - | - |
| Zhang et al., 2019 [21] | Retrospective | MRI | 26 CRC patients with 193 LM | One cohort | MATLAB (MATLAB R2011b, MathWorks, Inc., Natick, MA, USA) | Student’s t test or Mann–Whitney U test when not normally distributed. Multivariable logistic regression analysis |
| Lubner et al., 2015 [22] | Retrospective | CT | 77 CRLM patients | One cohort | TexRAD Ltd., (Somerset, UK) | Correlated using Cox proportional hazards models |
| Simpson et al., 2017 [23] | Retrospective | CT | 198 patients | One cohort | Scout Liver (Pathfinder Technologies Inc., Nashville, TN, USA) | Kaplan–Meier and Cox proportional hazards models |
| Ganeshan et al., 2007 [24] | Retrospective | CT | 27 patients | One cohort | TexRAD Ltd., (Somerset, UK) MATLAB (Mathworks Inc, Natick, MA, USA) | Cox regression analysis and the statistical significance of contingency tables was assessed using Fischer’s exact test. |
| Rahmim et al., 2019 [25] | Retrospective | FDG PET/CT | 52 CRLM patients | One cohort | Hermes Hybrid Viewer PDR and MATLAB | Kaplan–Meier and Cox proportional hazards models |
| Dercle et al., 2020 [26] | Retrospective | CT | 667 CRLM patients | Two cohorts. Randomly assigned (2:1) to training or validation sets. Predicted tumor sensitivity to treatment was measured using AUC in the validation sets of the four cohorts consisting of patients that were not used for training. | MATLAB (Mathworks, Natick, MA, USA) | Variance and v2 test were performed to compare categorical variables. Cox regression was used to investigate the effect of survival variables, and log-rank test was used to compare survival times of two groups. |
| Dohan et al., 2019 [27] | Multicenter prospective | CT | 491 CRLM patients treated by FOLFIRI and bevacizumab | Two cohorts. Training cohort in 120 patients, and validate cohort in 110 patients. External validation was performed in another cohort of 40 patients | TexRAD Ltd., (Somerset, UK) | Multivariable Cox, Kaplan–Meier and log-rank |
| Ravanelli et al., 2019 [28] | Retrospective | CT | 43 CRLM patients | Two cohorts. 23 treated with bevacizumab-containing chemotherapy (group A), and 20 with standard chemotherapy (group B) | MATLAB (Natick, MA, USA) | Multivariable logistic regression |
CT: computed tomography; MRI: magnetic resonance imaging; FDG PET/CT: fluorodeoxyglucose positron emission tomography/computed tomography; CRC: colorectal carcinoma; LM: liver metastases; CRCLM: colorectal carcinoma liver metastases; HCC: hepatocellular carcinoma; HHs: hepatic hemangiomas; FLLs: focal liver lesions; AUC: area under curve; ROC: receiver operating characteristic.