Table 3.
Radiomics and artificial intelligence (AI) studies in treatment response
| References | Imaging modality | Major end point | N (training + test cohorts) | Number of features in the model | Evaluation metric | Analytical method used | Study description | Validation type |
|---|---|---|---|---|---|---|---|---|
| Trebeschi et al. 2019 | CT | Response to immunotherapy | 133 + 70 | 2 | AUC = 0.83, P < 0.001 |
Machine learning | Identifying radiomic biomarkers for immunotherapy response | Internal validation with a separate data set |
| Sun et al. 2018 | CT | Response to immunotherapy | 135 + 256 | 8 | AUC = 0.76 | Machine learning | Identifying radiomic markers to assess tumor infiltrating CD8 cells and immunotherapy response | External validation |
| Tunali et al. 2019a | CT | Rapid disease progression | 228 | 4 radiomic + 4 clinical | AUC = 0.804 | Logistic regression | Identifying rapid disease progression phenotypes in patients treated with immunotherapy | Internal validation with bootstrapping |
| Tunali et al. 2019c | CT | Response to immunotherapy | 180 + 90 + 58 | 1 radiomic + 2 clinical | Log-rank P < 0.001 |
Classification and regression tree | Identifying survival risk groups for patients treated with immunotherapy | External validation |
| Mu et al. 2019a | FDG-PET/CT | Response to immunotherapy | 194 + 47 + 48 | 8 | AUC = 0.81 | LASSO | Predicting durable clinical benefit from immunotherapy | Internal validation with a separate data set |
| Cook et al. 2015 | FDG-PET | Response to Erlotinib | 47 | 1 delta radiomica | P = 0.01 | Wilcoxon signed-rank test | Identifying baseline radiomic biomarkers and delta-radiomic biomarkers of treatment response and OS of patients treated with EGFR TKI | No validation |
| Ravanelli et al. 2018 | CT | Response to Erlotinib | 50 | 5 | AUC = 0.8 | LASSO | Determining high- and low-risk groups of patients treated with EGFR TKI | Cross-validation and bootstrapping |
| Park et al. 2018 | FDG-PET/CT | Response to Erlotinib | 161 + 21 | 8 univariable features | C-index range = 0.630–0.669 | Harrell's C-index | Determining overall survival risk groups of patients treated with EGFR TKI | Internal validation with a separate data set |
| Coroller et al. 2017 | CT | pCR | 85 | 1 | AUC = 0.75, P = 0.01 |
Random forest | Predicting pathological response after neoadjuvant chemoradiation | No validation |
| Yu et al. 2018 | CT | Response to surgery or SABR | 147 + 295 | 2 | HR = 1.27, P < 2 × 10−16 |
Random survival forests | Predicting OS of stage I NSCLC patients | Internal validation with a separate data set |
| Mattonen et al. 2016 | CT | Local recurrence | 45 | 5 | FPR = 24.0%, FNR = 23.1% | Machine learning | Assess physician ability to detect timely local recurrence and to compare physician performance with a radiomics tool | No validation |
| Khorrami et al. 2019 | CT | OS and TTP | 72 + 53 | 7 | AUC = 0.77 | Minimum redundancy, maximum relevance | Discriminative ability of radiomic features on response to chemotherapy | Internal validation with a separate data set |
| Fave et al. 2017 | CT | Local recurrence | 107 | 1 delta radiomica | Log-rank P-value = 0.269, C-index = 0.558 |
Cox proportional hazards regression | Assessing radiation therapy response by using delta radiomics | No validation |
(HR) Hazard ratio, (OS) overall survival, (EGFR) Epidermal Growth Factor Receptor, (ANN) artificial neural network, (FPR) false positive rate, (FNR) false negative rate, (TTP) time-to-progression, (AUC) area under the curve, (pCR) pathological complete response, (SABR) stereotactic ablative radiotherapy, (LASSO) least absolute shrinkage and selection operator, (NSCLC) non-small-cell lung cancer.
aDelta-radiomics is the difference in radiomic features between multiple scans.