Table 5.
Summary of representative studies on 18F-FDG PET/CT radiomics-based machine learning analyses in abdominal tumors
| Authors | Years | Tumor type | Aim | Sample size | Constructed ML models | Core ML algorithm | Best ML model | Validation | Resultsa |
|---|---|---|---|---|---|---|---|---|---|
| Differentiating benign from malignant tumors | |||||||||
| Zhang et al. [102] | 2019 | Pancreatic tumor | AIP vs. PDAC | n = 251 |
CT radiomics-based model PET radiomics-based model Combined model |
SVM | Combined model | Internal validation (cross-validation) | Accuracy: 0.850 |
| Wei et al. [103] | 2023 | Pancreatic tumor | AIP vs. PDAC | n = 112 |
CT radiomics-based + PET radiomics-based model DL feature-based model Multidomain fusion model (radiomics + DL features) |
VGG11 DL algorithm | Multidomain fusion model | Internal validation (cross-validation) | Accuracy: 0.901 |
| Predicting tumor characteristics or stage | |||||||||
| Xing et al. [104] | 2021 | PDAC | Pathological grade | n = 149 |
CT radiomics-based model PET radiomics-based model Combined model |
XGB | Combined model | Training and validation cohorts | AUC: 0.921 |
| Jiang et al. [105] | 2022 | HCC or ICC | MVI | HCC: n = 76; ICC: n = 51 |
Clinical model CT radiomics-based model PET radiomics-based model Combined optimal PET and CT radiomics-based model Combined best clinical, PET radiomics-based, or CT radiomics-based model |
RF | Combined best clinical and PET feature-based model | Training and validation cohorts |
AUC for HCC: 0.88 AUC for ICC: 0.90 |
| Liu et al. [106] | 2021 | Gastric cancer | LNM | n = 185 |
CT radiomics-based model PET radiomics-based model Combined model |
Adaboost | Combined model | Training and validation cohorts | Accuracy: 0.852 |
| He et al. [107] | 2021 | Colorectal cancer | LNM | n = 199 | Combined CT + PET radiomics-based model | XGB | – | Training and validation cohorts | Accuracy: 0.7636 |
| Li et al. [108] | 2021 | Colorectal cancer | MSI | n = 173 | Combined clinical + CT radiomics-based + PET radiomics-based model alone | Adaboost | – | Training and validation cohorts | AUC: 0.828 |
| Predicting treatment response or survival | |||||||||
| Toyama et al. [109] | 2020 | Pancreatic cancer | 1-year survival after RT, CRT, or surgery | n = 161 | PET radiomics-based model alone | RF | – |
Internal validation (cross-validation) |
HR for GLZLM_GLNU: 2.0 |
| Liu et al. [110] | 2023 | Gastric cancer |
HER2 status Progression after surgery |
n = 90 |
Combined clinical + CT radiomics-based + PET radiomics-based model |
Adaboost | – | Training and validation cohorts |
Accuracy for HER2: 0.833 Accuracy for progression: 0.778 |
| Lv et al. [111] | 2022 | Colorectal cancer | Recurrence-free survival after surgery | n = 196 |
Clinical model CT radiomics-based model PET radiomics-based model Combined model |
RSF | Combined model | Training and validation cohorts |
C-index for all patients: 0.780 C-index for patients with stage III disease: 0.820 |
| Shen et al. [112] | 2020 | Rectal cancer | pCR after NCRT | n = 169 | PET radiomics-based model alone | RF | – | Internal validation | Accuracy: 0.953 |
| Agüloğlu et al. [113] | 2023 | Metastatic rectal cancer | 2-year OS | n = 62 | PET radiomics-based model alone | RF | – |
Internal validation (cross-validation) |
AUC: 0.843 |
AIP autoimmune pancreatitis, AUC area under the receiver operating characteristic curve, C-index concordance index, CRT chemoradiotherapy, DL deep learning, GLNU gray-level non-uniformity, GLZLM gray-level zone length matrix, HCC hepatocellular carcinoma, HER2 human epidermal growth factor receptor, HR hazard ratio, ICC intrahepatic cholangiocarcinoma, LNM lymph node metastasis, ML machine learning, MSI microsatellite instability, MVI microvascular invasion, NCRT neoadjuvant chemoradiotherapy, OS overall survival, pCR pathological complete response, PDAC pancreatic ductal adenocarcinoma, RF random forest, RSF random survival forest, RT radiotherapy, SVM support vector machine, XGB gradient tree boosting
aPerformance only presents the result of the best machine learning model