Clinical application of 18F-fluorodeoxyglucose positron emission tomography/computed tomography radiomics-based machine learning analyses in the field of oncology

Machine learning (ML) analyses using 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography (PET)/computed tomography (CT) radiomics features have been applied in the field of oncology. The current review aimed to summarize the current clinical articles about 18F-FDG PET/CT radiomics-based ML analyses to solve issues in classifying or constructing prediction models for several types of tumors. In these studies, lung and mediastinal tumors were the most commonly evaluated lesions, followed by lymphatic, abdominal, head and neck, breast, gynecological, and other types of tumors. Previous studies have commonly shown that 18F-FDG PET radiomics-based ML analysis has good performance in differentiating benign from malignant tumors, predicting tumor characteristics and stage, therapeutic response, and prognosis by examining significant differences in the area under the receiver operating characteristic curves, accuracies, or concordance indices (> 0.70). However, these studies have reported several ML algorithms. Moreover, different ML models have been applied for the same purpose. Thus, various procedures were used in 18F-FDG PET/CT radiomics-based ML analysis in oncology, and 18F-FDG PET/CT radiomics-based ML models, which are easy and universally applied in clinical practice, would be expected to be established.

The development of artificial intelligence (AI) is associated with relevant psychological, ethical, and medicolegal issues, which should be addressed before AI can be completely considered in patient management.However, the ultra-rapid analysis of large datasets is a major strength of AI in healthcare applications.In the field of medical imaging, AI has been significantly beneficial in predicting individual patient outcomes [7,8].Machine learning (ML) can resolve complex interactions among numerous variables to construct a prediction model as accurate as possible [9][10][11].The flexibility and scalability of ML are superior to those of conventional statistical approaches.Hence, ML is useful in several tasks including diagnosis and classification.
Recently, the ML or deep learning (DL) models using 18 F-FDG PET/CT radiomic features have been applied to resolve issues in classification (i.e., "benign or malignant tumor," "primary or metastatic tumor," "classification of histological subtypes," and "recurrence or non-recurrence") or to construct prediction models (i.e., "tumor characteristic," "tumor stage," or "survival") [12].The current review aimed to summarize the current clinical studies on 18 F-FDG PET/ CT radiomics-based ML analyses to address issues in classification or to construct prediction models for several types of tumors.

Literature search and screening
On April 20, 2023, we searched studies with the following terms in the title from PubMed: "PET/CT" and "radiomic" or "radiomics" and "machine learning." In total, 224 articles were identified during the initial search.The titles, abstracts, and texts were assessed to identify relevant articles.The inclusion criteria were as follows: (1) studies written in English, (2) original clinical studies about oncology, and (3) studies describing the application of the 18 F-FDG PET/CT radiomics-based ML approach for solving issues associated with classifying or constructing prediction models.The exclusion criteria were as follows: 1) reports only describing the CT radiomics-based ML approach, 2) studies using ML for image reconstruction or segmentation, 3) cohort studies with < 20 patients, and 4) review articles.Of 224 articles identified, 45 were review articles; hence, they were not included in the study.Among the remaining 179 original articles, 86 were excluded because of non-18 F-FDG tracer (n = 38), only CT-based radiomic ML analysis (n = 31), non-oncological disorders (n = 11), application of ML for image reconstruction or image segmentation (n = 4) and nonclinical studies (n = 2).Finally, 93 articles were included in the analysis, and all articles were published after 2018 (Fig. 1).

Difference between benign and malignant tumors and between primary and metastatic tumors
Pulmonary nodules are common clinical findings, and lung cancer frequently presents as a solitary pulmonary nodule (SPN) on diagnostic imaging at the early disease stage [13].SPNs are often incidentally detected.Thus, benign SPNs should be clinically differentiated from malignant SPNs.
Ren et al. [14] reported that the ML model with the least absolute shrinkage and selection operator (LASSO) regression algorithm using combined clinical data and PET-radiomics had a good diagnostic performance for distinguishing benign from malignant SPNs, with an area under the receiver operating characteristic curve (AUC) of 0.94.Zhou et al. [15] examined the ability of 18 F-FDG PET/CT radiomicsbased ML analysis in differentiating primary from metastatic lung lesions.Results showed that the ML model with the gradient boosting decision tree algorithm with PET-radiomics had the highest classification accuracy, with an AUC of 0.983.Some studies have found similar results [16][17][18][19] (Table 1).Thus, 18 F-FDG PET/CT radiomics-based ML analysis can have a great potential in characterizing SPNs.

Classification according to histological types
Due to the different histologic and biological characteristics of lung adenocarcinoma (ADC) and lung squamous cell carcinoma (SCC), their treatment regimen, prognosis, and relapse rate significantly vary [20,21].Thus, it is important to distinguish these two subtypes of non-small cell lung cancer (NSCLC) before treatment for appropriate clinical decision-making. 18F-FDG PET/CT radiomics-based ML analysis might improve the classification of ADC and SCC [22][23][24][25][26]. Zhao et al. [22] established combined ML models based on clinical characteristics (sex and smoking status), laboratory findings (carcinoembryonic antigen and squamous cell carcinoma antigen levels), and PET-radiomics to classify ADC and SCC.The support vector machine (SVM) algorithm accurately distinguished ADC from SCC, with an AUC of 0.876.This algorithm had a significantly better prediction performance than the clinical model (AUC:0.712,p = 0.037).Han et al. [23] examined the usefulness of PET radiomics-based ML/DL algorithms for obtaining differential diagnosis in patients with ADC and SCC.They reported that ML analyses with either the linear discriminant analysis (AUC: 0.863) or the SVM (AUC: 0.863) algorithm had optimal performance.Moreover, the VGG16 DL algorithm (AUC: 0.903) outperformed all conventional ML algorithms.Similar studies have successfully differentiated ADC from SCC [24][25][26] (Table 1).18 F-FDG PET/CT radiomics-based ML analysis can characterize histological subtypes in thymic epithelial tumors (TETs) [27,28].The ML model trained using 18 F-FDG PET radiomics and DL-based features with the logistic regression (LR) algorithm was proposed for predicting the histological subtypes of TETs [27].This model can accurately differentiate thymic cancer from thymoma, with an AUC of 0.90.

Prediction of tumor characteristics
Recently, the treatment options for NSCLC significantly improved with advancements in targeted therapies against mutated genes such as epidermal growth factor receptor (EGFR), kirsten rat sarcoma viral oncogene (KRAS), and anaplastic lymphoma kinase (ALK) [29,30].Moreover, immune checkpoint inhibitors targeting programmed cell death protein 1 (PD-1) or programmed death ligand 1 (PD-L1) are associated with better survival outcomes compared with conventional chemotherapy in patients with advancedstage NSCLC [31,32].Thus, in patients with NSCLC, gene mutations or the immune checkpoint status of tumors should be identified to determine the appropriate treatment strategy.
Several reports have examined the usefulness of 18 F-FDG PET/CT radiomics-based ML analysis for predicting gene mutation.Previous studies commonly showed that 18 F-FDG PET/CT radiomics-based ML analysis had a promising performance for predicting gene mutation [33][34][35][36][37][38][39] (Table 1).Gao et al. [33] constructed radiomics-based models based on 18 F-FDG PET/CT features using ML to predict EGFR mutation status in patients with lung ADC.Results showed that the ML model with the random forest (RF) algorithm using combined clinical data, CT-radiomics and PET-radiomics had the highest performance, with an AUC of 0.730.Chang et al. [34] revealed that the combined clinical data and PET/CT-based ML model with the LASSO regression algorithm is significantly advantageous in predicting ALK mutation status in patients with lung ADC compared with the clinical model (AUC:0.88 vs. 0.74, p < 0.001).Shiri et al. [35] reported that the ML model with the stochastic gradient descent algorithm using CT-radiomics and PETradiomics outperformed conventional methods (peak of standardized uptake value [SUVpeak]  F-FDG PET/CT radiomics-based ML model have a similar performance, with AUCs of 0.797-0.870[36][37][38][39]. Several studies have assessed the predictive ability of 18 F-FDG PET/CT radiomics-based ML analysis for immune checkpoint status in NSCLC [40][41][42].Lim et al. [40] predicted the PD-L1 expression level in patients with NCSLC using the 18 F-FDG PET/CT radiomics-based ML model.Results showed that the ML model with the Naïve Bayes algorithm using the top five features (CT_gray-level run length matrix [GLRLM]_long run high grey-level emphasis, CT_grey-level co-occurrence matrix [GLCM]_homogeneity, CT_mean Hounsfield unit, CT_GLRLM_long run emphasis, and PET_SUVmax) had the best predictive performance (AUC: 0.712).Mu et al. [41] developed a 18 F-FDG PET/CTbased DL model to evaluate PD-L1 status.Results showed that the deep learning score (DLS) could significantly distinguish PD-L1-positive from PD-L1-negative patients (AUC: 0.82).

Predicting tumor stage
The clinical outcome of NSCLC is directly related to its stage at diagnosis [43].Moreover, there were reports showing the usefulness of the 18 F-FDG PET/CT radiomicsbased ML method for predicting tumor stage in lung cancer [44][45][46].Wang et al. [44] reported that the ML model with the gradient tree boosting (XGB) ML algorithm using combined clinical data and PET/CT radiomics of the primary tumor and lymph node had the highest diagnostic performance in predicting lymph node metastasis (LNM) in NSCLC (AUC: 0.93).Moreover, this model had a great potential in predicting N2 stage NSCLC (AUC: 0.94).In addition, Laros et al. [45] reported that the combined PETradiomics of the primary tumor and lymph node had good performance in predicting LNM from NSCLC, with an accuracy of 0.88.
Zhao et al. [47] examined the ability of ML models trained using clinical data and 18 F-FDG PET/CT radiomics for predicting overall survival (OS) in patients with lung ADC who underwent surgery and received radiotherapy (RT), chemotherapy, or immunotherapy.The ensemble ML models, which were constructed with clinical data and 18 F-FDG PET/CT radiomic features, could predict the 3-and 4-year OS, with an AUC of 0.84 and 0.88, respectively.Huang et al. [48] showed that the convolutional neural networks (CNNs) trained by 18 F-FDG PET/CT had good performance in predicting OS in patients with malignant lung tumor who received RT, chemotherapy, or immunotherapy.To predict OS, the CNNs trained using clinical data and 18 F-FDG PET/CT radiomics with the random survival forest (RSF) ML model (concordance index [C-index]: 0.737) had a similar performance to CT alone (C-index: 0.730).However, it had a better performance than PET (C-index: 0.595) and clinical models (C-index: 0.595) alone.
Previous studies have assessed the ability of 18 F-FDG PET/CT radiomics-based ML models for predicting outcomes in not only patients with surgically treated NSCLC [49,50] but also those with nonsurgically treated NSCLC [51][52][53][54][55][56].Ahn et al. [49] used the 18 F-FDG PET/CT radiomics-based ML approach to predict disease recurrence in patients with NSCLC who underwent surgery.Results showed that the ML model with the RF algorithm had good performance for predicting recurrence, with an AUC of 0.956.Mu et al. [51] established the 18 F-FDG PET-based DLSs, which is useful for predicting EGFR mutation status (EGFR-DLS) (AUC: 0.81).EGFR-DLS was significantly and positively associated with a longer progression-free survival (PFS) in patients treated with EGFR-tyrosine kinase inhibitors (hazard ratio [HR]:0.24,p < 0.001).Mu et al. [52] reported that the 18 F-FDG PET/CT radiomicsbased ML model had a good AUC for predicting response to immune checkpoint inhibitors (0.81).Moreover, the constructed nomogram models (C-indices of 0.77 and 0.80 for predicting OS and PFS, respectively) had good performance in predicting prognosis.Similar studies have successfully predicted treatment responses or survival in patients with NSCLC [50,[53][54][55][56][57] (Table 1).
The 18 F-FDG PET/CT radiomics-based ML analysis has been applied to predict PFS in malignant pleural mesothelioma [58].This study showed the prognostic potential of the cox regression ML model established using specific PET radiomics-based on the principal component analysis for PFS with a C-index of 0.66.

Summary
Previous studies commonly showed that 18 F-FDG PET radiomics-based ML analysis had a high predictive performance for differentiating benign from malignant tumors, predicting tumor characteristics, staging tumors, and assessing treatment outcome or prognosis in lung or mediastinal tumors, with AUCs, accuracies, or C-indices of > 0.70.Thus, the 18 F-FDG PET radiomics-based ML analysis might play important roles in supporting clinicians in diagnostic and patient management including precision medicine for lung or mediastinal tumors.However, as shown in Table 1, previous studies have reported several ML processes including ML algorithms, and different ML models have been applied for the same purpose.

Differentiating benign and malignant tumors and predicting tumor characteristics
In head and neck tumors, 18 F-FDG PET/CT radiomicsbased ML analyses have been applied to differentiate benign from malignant tumors or to predict tumor characteristics.The following articles have reported about differentiating benign from malignant tumors.
In thyroid incidentalomas, distinguishing benign from malignant tumors based on SUVmax on 18 F-FDG PET/CT is challenging due to a significant overlap between these lesions [59].Aksu et al. [60] reported that the ML model with the RF algorithm had a better performance in differentiating benign from malignant thyroid incidentalomas based on SUVmax (AUC: 0.849 vs. 0.758).
The assessment of human papillomavirus (HPV) status plays an important role in treatment planning for oropharyngeal cancer [61].Haider et al. [62] showed that the AUC of combined tumor and lymph node PET/CT radiomics-based ML model with the XGB algorithm for predicting HPV status in oropharyngeal cancer was 0.83.

Predicting treatment response or survival
Previous studies have reported the predictive ability of 18 F-FDG PET/CT radiomics-based ML analysis for treatment outcomes in head and neck cancers [63][64][65][66][67][68][69][70][71] (Table 2).Haider et al. [63] showed that the ML model with the RSF algorithm using clinical and pretreatment 18 F-FDG PET/CT radiomics had good predictive performance for locoregional progression in patients with HPV-associated oropharyngeal cancer who received RT (C-index: 0.76).In hypopharyngeal cancers, the ML model with the LR algorithm constructed based on UICC stage, T and N stage, and pretreatment 18 F-FDG PET-radiomics with GLCM_entropy and GLRLM_ run length non-uniformity (RLNU) is a significant predictor of PFS (HR:3.22,p = 0.045) [64].
Previous studies have reported the usefulness of intratreatment 18 F-FDG PET/CT radiomics-based ML analysis for outcome prediction in head and neck cancers.Lafata et al. [65] showed that the unsupervised clustering of intratreatment 18 F-FDG PET/CT radiomics, which were obtained 2 weeks after RT (at a dose of 20 Gy), was significantly associated with recurrence-free survival (HR:2.69,p = 0.04) in patients with oropharyngeal cancer who received definitive RT.Moreover, a previous study assessed the ability of ML analysis using the combined 18 F-FDG PET radiomics and genomic data for predicting 3-year OS in head and neck cancers (AUC: 0.75) [66].Similar studies have successfully predicted prognosis in head and neck cancer [67][68][69][70][71] (Table 2).

Summary
Previous studies revealed that 18 F-FDG PET/CT radiomicsbased ML analysis had good predictive performances for predicting treatment outcome or prognosis, with AUCs or C-indices of > 0.70, in head and neck tumors.Thus, 18 F-FDG PET/CT radiomics-based ML analysis might be expected to be an important tool for patient management in head and neck tumors.However, several ML processing approaches have also been discussed (Table 2).

Differentiating benign from malignant tumors and primary from metastatic tumors or classifying tumors according to histological types
The conventional semi-quantitative 18 F-FDG PET parameters such as SUVmax, MTV, and total lesion glycolysis (TLG) are useful biomarkers for characterizing malignant lymphoma [72,73].However, the ability of these parameters in identifying tumor heterogeneity, which ultimately contributes to tumor aggressiveness and poor prognosis, remains limited [74].Recently, 18 F-FDG PET/CT radiomics-based ML analysis has been applied to overcome these issues [75].Previous studies have revealed that 18 F-FDG PET/CT radiomics-based ML analysis is useful in not only classifying tumors based on histological subtypes but also differentiating malignant lymphoma from other diseases [76][77][78][79][80].
Abenavoli et al. [76] showed that the ML model with the RF algorithm using PET-radiomics had a better performance in differentiating diffuse large B-cell lymphoma (DLBCL) from Hodgkin's lymphoma (HD) based on SUVmax (AUC: 0.87 vs. 0.78).de Jesus et al. [77] reported that the ML model with the gradient boosting algorithm using PET/CT radiomics had a significantly higher AUC in distinguishing DLBCL and follicular lymphoma according to SUVmax (AUC:0.86 vs. 0.79, p < 0.01).Lovinfosse et al. [78] also showed that the ML model with the RF algorithm using clinical data and PET-radiomics had good performance in differentiating DLBCL from HD, with an AUC of 0.95.Further, the authors showed that the constructed ML model with the RF algorithm had good performance in differentiating malignant lymphoma and sarcoidosis, with an AUC of 0.94.Yang et al. [79] revealed that the ML model with the SVM algorithm constructed according to combined CNN-based features and PET-radiomics had a great potential in distinguishing malignant lymphoma from enlarged metastatic cervical lymph nodes (AUC: 0.948).

Predicting treatment response or survival
For the treatment assessment of malignant lymphoma, the visual assessment of the Deauville score (DC) has been a useful 18 F-FDG PET/CT criterion: DC1-DC3, complete metabolic response; DC4 and DC5, incomplete metabolic response [81][82][83].However, there might be difficulties in predicting treatment outcomes based on DC alone because of the inter-or intra-variability of DC definition.Thus, 18 F-FDG PET/CT radiomics-based ML analysis can be a novel approach for predicting treatment outcomes in malignant lymphoma.
Frood et al. [84] examined the ability of pretreatment 18 F-FDG PET/CT radiomics-based ML analysis for predicting recurrence after DLBCL treatment.Results showed that the ML model with the ridge regression algorithm using combined clinical and PET-radiomics had good performance, with an AUC of 0.73.Cui et al. [85] assessed the potential of the 18 F-FDG PET/CT radiomics-based ML approach for identifying patients with DLBCL who are at high risk for progression or relapse after receiving first-line therapy.Results showed that the ML model with the RF algorithm using clinical data, baseline, end-of-treatment and delta PET-radiomics features was a significant predictor of PFS (C-index: 0.853).By contrast, 18 F-FDG PET/CT radiomicsbased ML analysis was found to be useful for predicting recurrence after HD treatment [86,87].Frood et al. [86] showed that the ML model with the ridge regression algorithm using combined clinical data and PET-radiomics had good predictive performance, with an AUC of 0.81.Similar studies have successfully predicted treatment responses or survival in malignant lymphoma [87][88][89][90][91] (Table 3).

Summary
Previous studies have shown that 18 F-FDG PET/CT radiomics-based ML analysis is useful in not only differentiating but also predicting treatment outcome or prognosis in patients with malignant lymphomas.Each best ML model had good predictive performance, with AUCs or C-indices of > 0.70 (Table 3).Thus, it might be expected to promote the translation of 18 F-FDG PET/CT radiomics-based ML analysis into clinical practice in the field of lymphatic tumors.However, the articles included in this review showed heterogeneity among various ML approaches.

Differentiating benign from malignant tumors and predicting tumor characteristics or stage
Several studies have examined the clinical potential of 18 F-FDG PET/CT radiomics-based ML analyses in differentiating benign from malignant tumors and predicting tumor characteristics or stage in breast cancer [92][93][94][95][96].
Eifer et al. [92] showed that ML analyses with the k-nearest neighbors (kNN) algorithm using CT-radiomics and PET-radiomics had good performance in differentiating LNM from breast cancer from post-COVID-19 vaccineassociated axillary lymphadenopathy, with an AUC of 0.98.
An accurate assessment of both hormone receptor status and human EGFR 2 (HER2) status is important for treatment planning in breast cancer [97,98].Moreover, an accurate pretreatment assessment of axillary lymph node is essential in managing breast cancer [99].Chen et al. [93] showed that the constructed ML model with the XGB algorithm based on PET/CTmean radiomics had good predictive ability for HER2 status in breast cancer (AUC: 0.76).In addition, Song [94] reported that the constructed ML model with the XGB algorithm based on PET/CT radiomics had good performance for predicting axillary LNM in patients with breast cancer (AUC: 0.890).A similar study has successfully predicted hormone status in breast cancer [95] (Table 4).

Predicting treatment response or survival
Two studies have examined the ability of 18 F-FDG PET/ CT radiomics-based ML analysis for predicting treatment outcome in breast cancer [100,101].Li et al. [100] assessed the usefulness of 18 F-FDG PET/CT radiomics-based ML analysis for predicting pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in breast cancer.Results showed that the diagnostic accuracy of the ML model with the RF algorithm constructed based on patient age and PET/CT radiomics increased compared with that of the ML model with the RF algorithm constructed according to PET/CT radiomics only (0.800 vs. 0.767).The authors hypothesized that the finding can be attributed to the fact that younger patients had a higher pCR rate than older ones.Gómez et al. [101] assessed the predictive ability of 18 F-FDG PET/CT radiomics-based ML analysis for metabolic response after metastatic breast cancer treatment.Results showed that the ML model with the LASSO + SVM algorithm using combined clinical data and PET-radiomics had good performance, with an AUC of 0.82.

Summary
In breast tumors, each best ML model had good predictive performance for differentiating benign from malignant tumors and for predicting tumor characteristics and stage and treatment outcome, with AUCs or accuracies of > 0.70 (Table 4).The heterogeneity of ML approaches was also noted in the reported studies.
Although there have been several studies that have explored the usefulness of 18 F-FDG PET/CT radiomicsbased ML analysis associated with breast tumors, it might be expected in the 18 F-FDG PET/CT radiomics-based ML analysis to be a novel tool to patient management for breast tumors.

Differentiating benign from malignant tumors and predicting tumor characteristics or stage
In abdominal tumors, the usefulness of 18 F-FDG PET/CT radiomics-based ML analyses in differentiating benign and malignant tumors and predicting tumor characteristics or stage has been evaluated [102][103][104][105][106][107][108] (Table 5).In pancreatic tumors, the ML model with the SVM algorithm using CT-radiomics and PET-radiomics has been a useful tool for differentiating autoimmune pancreatitis (AIP) and pancreatic ductal adenocarcinoma (PDAC), with an accuracy of 0.850 [102].Moreover, this group established the multidomain fusion DL model using CT-radiomics, PET-radiomics, and DL features for differentiating AIP from PDAC [103].Results showed that the accuracy of this DL model improved (0.901) compared with that of the formerly published ML model [102].Xing et al. [104] assessed the ability of 18 F-FDG PET/CT radiomics-based ML analysis for predicting the pathological grade of PDAC.Results showed that the ML model with the XGB algorithm using the combined CT-radiomics and PET-radiomics (AUC: 0.921) was better in predicting the pathological grade of PDAC than the CT-radiomics alone (AUC: 0.817) or the PET radiomicsbased model alone (AUC: 0.771).
In liver tumors, Jiang et al. [105] assessed the usefulness of 18 F-FDG PET radiomics-based ML analysis for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC).Results showed that the constructed ML model with the RF algorithm using PET-radiomics and clinical features (cancer antigen 19-9 level or tumor stage) was useful for predicting MVI in either HCC (AUC: 0.88) or ICC (AUC: 0.90) [105].
Liu et al. [106] constructed a useful ML model with the Adaboost algorithm using CT-radiomics and PET-radiomics for predicting LNM in gastric cancer with an accuracy of 0.852.This model detected some metastatic lymph nodes that were missed on contrast-enhanced CT scan (19.6%).Thus, the constructed ML model might offer a potentially useful adjunct to the current staging approaches for gastric cancer.He et al. [107] showed that the ML model with the XGB algorithm using CT-radiomics and PET-radiomics was successful in classifying regional LNM from colorectal cancer, with an accuracy of 0.7636.This ML model was better in predicting LNM than lymph node status, as described in clinical 18 F-FDG PET/CT scan reports (accuracy: 0.7091).Li et al. [108] reported that 18 F-FDG PET/CT radiomics-based ML analysis was useful for predicting the microsatellite instability (MSI) status, which is an essential prognostic factor of colorectal cancer.Results showed that the constructed ML model with the Adaboost algorithm using two selected radiomic features (PET-Skewness and CT-RoomMeanSquared) had good predictive performance for MSI, with an AUC of 0.828.

Predicting treatment response or survival
Several reports examined the usefulness of 18 F-FDG PET/ CT radiomics-based ML analyses for predicting treatment outcome in abdominal tumors [109][110][111][112][113].These studies showed that the 18 F-FDG PET/CT radiomics-based ML analyses were the power tools for predicting treatment response or prognosis.
Toyama et al. [109] revealed that PET-radiomics with gray-level zone length matrix (GLZLM)_gray-level nonuniformity (GLNU) was the most important feature on the ML model with the RF algorithm for predicting 1-year survival in pancreatic cancer, and multivariate analysis with Cox hazard regression revealed GLZLM_GLNU as the only statistically significant PET-radiomics for predicting 1-year survival (HR:2.0,p = 0.0094).Liu et al. [110] constructed the ML model with the Adaboost algorithm using clinical data, CT-radiomics and PET-radiomics for predicting HER2 expression status or disease progression in gastric cancer.The predictive accuracies of constructed ML model for HER2 expression status and disease progression were 83.3% and 77.8%, respectively.Lv et al. [111] developed the ML mode with the RSF algorithm using clinical data, CT-radiomics and PET-radiomics to predict recurrence-free survival in patients with colorectal cancer who received surgery, and revealed that the constructed ML model had good performance in predicting the prognosis (C-index for all patients, 0.780; C-index for stage III patients, 0.820).Shen et al. [112] constructed the ML model with the RF algorithm using PET-radiomics for predicting pCR after neoadjuvant chemoradiotherapy (CRT) in rectal cancer, and this ML model showed high predictive performance with an accuracy of 0.953.Moreover, the ability of ML model with the RF algorithm using PET-radiomics for predicting 2-year OS has also been reported in metastatic rectal cancer (2-year OS; AUC:0.843)[113].

Summary
In abdominal tumors, each best 18 F-FDG PET radiomicsbased ML model had good predictive performance for differentiating benign and malignant tumors, predicting tumor characteristics, staging tumors, or assessing treatment outcome with AUCs, accuracies, or C-indices of > 0.70 (Table 5).The application of 18 F-FDG PET radiomics-based ML analyses might be especially expected in the field of gastrointestinal cancers.

Predicting tumor stage
The expression of some protein molecules such as cyclooxygenase-2 (COX-2) is associated with LNM and lymphovascular space invasion (LVSI) in cervical cancer [114,115].Tumor budding (TB) is defined as a single neoplastic cell or cell cluster of up to four neoplastic cells at the invasive front of the tumor or within the tumor mass (intratumoral budding) [116].Moreover, TB is associated with LNM, LVSI, and prognosis in cervical cancer [117].Some investigators applied the 18 F-FDG PET/CT radiomics-based ML models for predicting not only LNM or LVSI but also the expression of COX-2 or TB status in cervical cancer [118][119][120][121] (Table 6).Lucia et al. [118] developed the ML model with the neural network algorithm using combat harmonized PETradiomics acquired from the different PET scanners (analog and digital PET) for predicting para-aortic LNM in cervical cancer.Results showed that the constructed ML model had an extremely high predictive ability, with an AUC of 0.96.Zhang et al. [119] showed that the constructed ML model with the LR algorithm using the PET-radiomics scores established using the LASSO regression had good predictive performance for not only pelvic LNM (AUC: 0.817) but also the expression of COX-2 (AUC: 0.814) in cervical cancers.Li et al. [120] revealed that the ML model with the LR algorithm using the PET-radiomics scores constructed using the LASSO regression had good predictive performance for LVSI in cervical cancer, with an AUC of 0.806.Chong et al. [121] showed that the constructed ML model with the SVM algorithm using conventional parameters (SUVmax, MTV, and TLG) and selected 29 PET-radiomics using the LASSO regression algorithm had good predictive performance for intratumoral budding in cervical cancer (AUC: 0.784).

Predicting treatment response or survival
A few reports have addressed the efficacy of 18 F-FDG PET/ CT radiomics-based ML analysis for predicting treatment outcomes or prognosis in cervical or endometrial cancer [122][123][124] (Table 6).
Ferreira et al. [122] showed that the ML model with the RF algorithm using clinical data and PET-radiomics had good performance for predicting disease-free survival in patients with advanced-stage cervical cancer who received CRT (AUC: 0.78).Another study revealed that the ML model with the Naïve Bayes algorithm constructed based on FIGO stage and four pretreatment PET-radiomics features (including surface area, MTV, GLRLM_RLNU, and GLRLM_GLNU) was a significant predictor of PFS (HR:6.89,p = 0.003) in patients with cervical cancer who underwent surgery and/or received CRT or chemotherapy [123].In endometrial cancers, the ML model with the kNN algorithm established using combined clinical data and PETradiomics has been useful for predicting disease progression, with an AUC of 0.890 [124].In this study, coarseness, which was the best PET-radiomics feature, was considered a significant and independent factor of PFS (HR:0.65,p = 0.003) and OS (HR:0.52,p < 0.001) in the multivariate Cox regression analysis.

Summary
In cervical or endometrial cancers, each best ML model had good predictive performance for predicting tumor stage with an AUC or accuracy of > 0.70.Moreover, the best ML model or best PET-radiomics feature is a significant predictor of survival, and the heterogenous ML approaches were also observed among the reported studies.Although there are not so many reports that have explored the usefulness of 18 F-FDG PET/CT radiomics-based ML analysis associated with gynecological tumors, the 18 F-FDG PET/CT radiomics-based ML analysis might provide useful information about patient management with gynecological tumors for clinicians.

Clinical application of 18 F-FDG PET/CT radiomics-based ML analyses in other tumors
In hematological malignancies including multiple myeloma and acute leukemia, 18 F-FDG PET/CT radiomics-based ML analyses have been applied to identify skeletal metastases, predict diffuse infiltration in the bone marrow, or predict prognosis [125][126][127][128] (Table 7).
Mannam et al. [125] showed that the ML model with the multilayer perceptron algorithm established based on CT-radiomics and PET-radiomics had good classification accuracy between multiple myeloma and skeletal metastases, with an AUC of 0.9538.Mesguich et al. [126] developed an ML model with the RF algorithm using five PET/ CT radiomics for predicting diffuse infiltration in the bone marrow in multiple myeloma.Results showed that the constructed ML model had an extremely high predictive ability, with an AUC of 0.90.Further, the ML model with the RF algorithm using CT-radiomics and PET-radiomics had good performance in predicting bone marrow involvement in acute leukemia [127].The diagnostic accuracy of this model was significantly higher than that of visual analysis (0.886 vs. 0.686, p = 0.041).Ni et al. [128] evaluated the ability of 18 F-FDG PET/CT radiomics-based ML analysis for predicting PFS after multiple myeloma treatment.Results showed that the ML model with the LASSO + cox regression algorithm trained using the combined clinical and PET/ CT radiomics-based model had a higher predictive performance (C-index: 0.698) than the ML model with clinical data (C-index: 0.563) or PET/CT radiomics-based model (C-index: 0.651) alone.The mistosis-karyorrhexis index (MKI) status is an independent prognostic factor of neuroblastoma [129].Feng et al. [130] developed the 18 F-FDG PET/CT radiomics-based ML model for predicting MKI status in neuroblastoma.The constructed ML model with the XGB algorithm using PET/ CT radiomics had an extremely high predictive ability, with an AUC of 0.951.Thus, the ML model can be used to noninvasively predict MKI status in pediatric neuroblastoma.Further, it is a significantly effective tool for the long-term management of pediatric neuroblastoma.

Conclusion
The efficacy of 18 F-FDG PET/CT radiomics-based ML analyses in various tumors was investigated.The number of studies about this topic has been increasing after 2018.The 18 F-FDG PET/CT radiomics-based ML analyses might be expected to be important tools for patient management in several types of tumors.However, previous studies have reported numerous ML procedures including the use of algorithms, and different ML models have been applied for the same purpose.Thus, various approaches are used to perform 18 F-FDG PET/CT radiomics-based ML analysis in oncology.Moreover, 18 F-FDG PET/CT radiomics-based ML models, which can be easily and universally applied in clinical practice, should be established.

Fig. 1 Excluded articles, n = 86 Non- 18 F
Fig. 1 Flowchart of study retrieval via literature search and inclusion in the analysis

Table 2
Summary of representative studies on 18 F-FDG PET/CT radiomics-based machine learning analyses in head and neck tumors

Table 3
Summary of representative studies on 18 F-FDG PET/CT radiomics-based machine learning analyses in lymphatic tumors

benign from malignant tumors and primary from metastatic tumors or classifying tumors according to pathological subtypes
a Performance only presents the result of the best machine learning model

benign from malignant tumors and predicting tumor characteristics or stage
a Performance only presents the result of the best machine learning model

Table 6
Summary of representative studies on 18 F-FDG PET/CT radiomics-based machine learning analyses in gynecological tumors

Table 7
Summary of representative studies on 18 F-FDG PET/CT radiomics-based machine learning analyses in other types of tumors AUC area under the receiver operating characteristic curve, C-index concordance index, LASSO least absolute shrinkage and selection operator algorithm, MKI mitosis-karyorrhexis index, ML