Table 1.
Summary of studies using PET-based radiomics for lung cancer
| Author | Imaging | Aim | Cohort size | Study population | Validation (Available or Not Available) | Outcome comparison | Results |
|---|---|---|---|---|---|---|---|
| Diagnosis and typing | |||||||
| Hu et al. (98) | 18F-FDG PET/CT | Distinguishing solitary ADC from tuberculosis | n = 235 (163 for training and 72 for validation) | ADC (n = 131) or tuberculosis (n = 104) | A | discrimination between ADC and tuberculosis | The AUC of the RF model was significantly higher than that of the clinical model and was slightly lower than that of the combined complex model |
| Zhang et al. (99) | 18F-FDG PET | Distinguishing tuberculosis nodules from lung cancer | n = 174 | tuberculosis nodules (n = 77) or lung cancer (n = 97) | NA | discrimination between tuberculosis nodules and lung cancer | The integrated model was found to be the best classification model |
| Zhang et al. (100) | 18F-FDG PET | Distinguishing ADC from SCC | n = 255(70% for training/validation and 30% for testing) | NSCLC | A | discrimination between ADC and SCC | the logistic regression classifier exhibited the most effective performance |
| Ji et al. (103) | 18F-FDG PET | Distinguishing ADC from SCC in different stages | n = 416 (253 for training and 163 validation) | stage I to III NSCLC patients diagnosed with ADC or SCC | A | discrimination between ADC and SCC | The AUCs of RF model for I to III stage in both the training and validation cohorts were good and the radiomic-clinical nomogram outperformed with higher AUCs |
| Zheng et al. (104) | 18F-FDG PET | Distinguishing benign and malignant SPN | n = 190 (70% for training/validation and 30% for testing) | SPN | A | discrimination between benign and malignant SPN | The combined effect is superior to qualitative diagnoses with CT or PET radiomics models alone. |
| Salihoğlu et al. (105) | 18F-FDG PET/CT | distinguishing between benign and malignant SPN | n = 48 (70% for training/validation and 30% for testing) | SPN | A | discrimination between benign and malignant SPN | The models provided reasonable performance for the differential diagnosis of SPNs (AUCs ~ 0.81) |
| Wang et al. (107) | 18F-FDG PET/CT | predicting for LVI of NSCLC | n = 148 (70% for training/validation and 30% for testing) | NSCLC | A | Predicting LVI in NSCLC patients | The integrated model was found to be the best classification model |
| Zheng et al. (108) | 18F FDG-PET/CT | predicting for brain metastasis of NSCLC | n = 203 (70% for training/validation and 30% for testing) | NSCLC | A | Predicting brain metastasis in NSCLC patients | The C-indices of the RF model in the training, internal validation, and external validation cohorts were 0.911, 0.825 and 0.800, respectively. |
| Jiang et al. (111) | 18F FDG-PET/CT | assessing PD-L1 expression status in NSCLC | n = 399 (2/3 for training and 1/3 for validation and model evaluation) | stage I-IV NSCLC | A | PD-L1 (SP142) and PD-L1 (28 − 8) expression level over 1% and over 50% prediction | Models based on CT-, PET/CT derived features anticipate PD-L1 expression status relatively accurate, while the CT-based model was superior |
| Zhang et al. (112) | 18F-FDG PET/CT | assessing PD-L1 expression status in NSCLC | n = 58 | NSCLC | NA | PD-L1 expression | Heterogeneity-related 18F-FDG PET and CT radiomic features, GLRLM_LGRE and GLZLM_SZE, could predict PD-L1 expression |
| Li et al. (113) | 18F-FDG PET/CT | assessing PD-L1 expression status in NSCLC | n = 136 (70% for training/validation and 30% for testing) | NSCLC | A | PD-L1 expression | the AUC of the fusion model was also higher than that of the RF model and the deep learning model |
| Liu et al. (120) | 18F-FDG PET/CT | identifying the specific EGFR mutation subtypes in ADC | n = 148 (111 for training and 37 for testing) | ADC | NA | specific EGFR mutation subtypes including EGFR-19-MT amd EGFR-21-MT) | The predictive features achieved AUCs of 0.77 for EGFR-19-MT, 0.92 for EGFR-21-MT and 0.87 for the combined EGFR mutation positivity |
| Li et al. (121) | 18F-FDG PET/CT | identifying the EGFR mutation status in NSCLC | n = 115 | NSCLC | NA | recognition of EGFR mutation | PET/CT based RFs achieved an AUC of 0.805 for discriminating between EGFR-MT and EGFR-WT |
| Zhao et al. (122) | 18F-FDG PET/CT | identifying the EGFR mutation status in ADC | n = 88 (65 for training and 23 for validation) | ADC | A | recognition of EGFR mutation | The model based on RFs combined with clinical factors achieved best discrimative performance with a AUC of 0.864 |
| Yang et al. (123) | 18F-FDG PET/CT | identifying the EGFR mutation status and specific subtypes and predicting the survival benefit of targeted TKIs therapy in NSCLC | n = 313 (70% for training and 30% for validation) | NSCLC | A | specific EGFR mutation subtypes including EGFR-19-MT amd EGFR-21-MT; OS; PFS | Radiomics models exhibited excellent ability to distinguish between EGFR-WT, EGFR-19-MT and EGFR-21-MT; the integrated nomogram was superior to the clinical nomogram and the radiomics nomogram, with C-indexes of 0.80 in the training set and 0.83 in the validation set |
| Yang et al. (124) | 18F-FDG PET/CT | Identifying the EGFR mutation status in ADC | n = 174 (139 for training and 35 for validation) | ADC | A | recognition of EGFR mutation; OS | The model achieved AUC of 0.77 in mutant/wild-type model and of 0.82 in 19/21 mutation site model; the multivariate CPH model achieved a C-index of 0.757 |
| Wang et al.(128) | 18F-FDG PET/CT | Identifying KRAS mutation status in NSCLC | n = 180 (180 for training and 78 for validation) | NSCLC | A | recognition of KRAS mutation | Integrating EGFR mutation information into the PET/CT RS model elevated the AUC, sensitivity, specificity, and accuracy. |
| Bourbonne et al.(129) | 18F-FDG PET/CT | Identifying KEAP1/NFE2L2 mutation status in NSCLC | n = 432 (158 for training and 274 for validation) | NSCLC | A | recognition of KEAP1/NFE2L2 mutation | The model achieved AUC of 0.8 in mutation prediction in the testing cohort and a hazard ratio of 2.61 in LR risk stratification. |
| Sanduleanu et al. (134) | 18FDG- PET/CT | Identifying hypoxia | n = 808 | patients with solid tumors | NA | recognition of hypoxic sites | Both disease-agnostic and lung-specific models achieved reasonable AUCs |
| Prognosis and efficacy evaluation | |||||||
| Li et al. (138) | 18F-FDG PET/CT | predicting the death and recurrent risk in ADC | n = 752 (including 4 gene expression datasets and 2 18F-FDG PET image datasets) | patients with ADC | NA | OS and RFS | The radiomic signature reflecting biological processes in tumors was significantly associated with patients’ OS and RFS (OS: log-rank P = 0.0006; RFS: log-rank P = 0.0013) |
| Chen et al. (139) | 18F-FDG PET/CT | predicting survival in ADC patients receiving targeted TKI treatment | n = 51 | stage III-IV ADC patients receiving targeted TKI treatment | NA | OS and PFS | A scoring system combining PET radiomics with clinical risk factors improved survival stratification |
| Yang et al. (144) | 18F-FDG PET/CT | predicting pCR to neoadjuvant chemoimmunotherapy in NSCLC | n = 185 | NSCLC | A | pathological complete response | The integrated model was found to be the best classification model |
| Nemoto et al. (146) | 18F-FDG PET/CT | predicting recurrence after SBRT | n = 82 | NSCLC | NA | local recurrence, regional lymph node metastasis, and distant metastasis | the model combining PET imaging features and SVM would be useful in predicting local and regional lymph node recurrence |
| Krarup et al. (148) | 18F-FDG PET/CT | predicting survival | n = 233 | NSCLC patients receiving definitive chemoradiotherapy | NA | PFS | The pre-selected RFs were insignificant in predicting PFS in combination with GTV, clinical stage and histology |
| Kirienko et al. (150) | 18F-FDG PET/CT | predicting DFS for patients undergoing surgery in NSCLC | n = 295 | NSCLC patients diagnosed with ADC or SCC | NA | DFS | The Cox models based on CT, PET, and PET/CT RFs achieved AUCs of 0.75, 0.68, and 0.68, respectively |
| Ouyang et al. (155) | 18F-FDG PET/CT | identifying metastatic LNs from the hypermetabolic mediastinal-hilar LNs in NSCLC | LN = 288 (159 LNs for training and 129 LNs for validation) | NSCLC patients with hypermetabolic LNs | A | recognition of metastatic LNs | PET/CT based model achieved the optimal AUC of 0.874 |
| Sepehri et al. (156) | 18F-FDG PET/CT | evaluating the potential benefit of combining different algorithms into a consensus for survival prediction | n = 138 (87 for training and validation and 51 for testing) | stage II and III NSCLC receiving curative (chemo) radiotherapy | A | median OS or OS shorter than 6 months | A consensus of machine learning algorithms could improve prognostic performance |
ADC, lung adenocarcinoma; AUC, area under the curve; RF, radiomics feature; NSCLC, non-small cell lung cancer; SBRT, stereotactic body radiotherapy; SCC, squamous cell carcinoma; SPN, solitary pulmonary nodules; EGFR, epidermal growth factor receptor; LVI, lymphovascular invasion; MT, mutant type; WT, wild type; OS, overall survival; pCR, pathological complete response; PFS, progression-free survival; C-index, concordance index; CPH, Cox proportional hazard; RFS, recurrence-free survival; GTV, gross tumor volume; DFS, disease-free survival; LN, lymph node; A, available; NA, not available