The resolution of soft tissue on MRI is significantly superior to that of CT and PET-CT and can effectively show the range of parapharyngeal space, skull base, and intracranial tumors. It is the gold standard for the evaluation of NPC (5–7). NPC is prone to distant metastasis; therefore, PET-CT is now used to evaluate distant metastasis by providing systemic anatomical and metabolic information. The National Comprehensive Cancer Network guidelines (23) recommend MRI and PET-CT for the diagnosis of NPC. The TNM staging system (based on tumor size, regional lymph node involvement, and distant metastases) that is currently used to guide the diagnosis and treatment of NPC is regarded as the gold standard. However, during diagnosis and treatment, patients at similar stages have exhibited different treatment responses, which may be due to the internal heterogeneity of the tumor. Thus, Zhu et al. (24) developed a radiomics model that combined the features extracted from MRI data with clinical information to analyze the survival subgroups of early NPC (validation group C-index: 0.814), compared it with the T (C-index: 0.803) and TNM staging systems (C-index: 0.765), and concluded that performance of the radiomics model was superior to the TNM staging system. This may have a significant impact on individualized diagnoses, treatments, and prognoses of NPC in the future.

Compared with CT, MRI has a higher resolution of soft tissue, which is advantageous for imaging NPCs. Research on the application of CT combined with radiomics to the diagnosis and staging of disease remains limited. PET-MRI combines the metabolic characteristics of PET and the high-resolution characteristics of MRI (6). Compared with MRI, the increased [¹⁸F]-fluorodeoxyglucose (FDG) uptake of PET-MRI can better show the subtle changes in local lesions, and it can also provide a more suitable anatomical reference than can PET-CT (5). Chan et al. (25) found that the sensitivity, specificity, and accuracy for the diagnosis of primary tumors of head and neck MRI were 94.2%, 90.9%, and 99.5% (p = 0.75), respectively, 99.6%, 98.3%, and 99.2% (p = 0.92) for head and neck [¹⁸F]-FDG PET-CT, respectively, and 98.2%, 96.3%, and 99.3% (p = 0.87) for [¹⁸F]-FDG PET-MRI, respectively. The positive predictive value of PET-MRI in the diagnosis of distant metastasis (93.1%) is higher than that of MRI and PET-CT (78.8% and 83.3%, respectively). This was a prospective study that suggested that this imaging method has better diagnostic capabilities for nasopharyngeal cancer and may play an important role in the diagnosis and treatment of NPC in the future.

In a study of PET-MRI combined with radiomics, Feng et al. (26) developed a radiomics model of FDG PET-MRI and reported areas under the curve (AUC) of the training group based on T2-weighted imaging and PET models of 0.85 and 0.84, respectively, and those of the validation group of 0.83 and 0.82, respectively, which offers great promise for the clinical staging of NPCs. In terms of internal heterogeneity of tumors, Akram et al. (27) showed that the imaging features Neighboring Gray Tone Difference Matrix-busyness extracted from MRI data before and after treatment may reflect differences between recurrent and non-recurrent areas in tumors; moreover, they demonstrated the potential of radiomics in the identification of radiation resistance in tumors before treatment to select dose increments.

Radiomics is advantageous not only for the diagnosis of diseases but also for differential diagnoses. The clinical manifestations and medical images of radiation-induced osteonecrosis and bone metastasis of the cervical spine are similar (28). Furthermore, the two conditions require different treatment methods and thus, require differentiation before treatment. The AUCs of MRI-based radiomics nomogram training and validation groups have been reported to be 0.725 and 0.720, respectively (28). Although CT and MRI are not applicable for differentiating between tumor recurrence and inflammation (29), PET-CT can distinguish between these two conditions; however, the high uptake of inflammation can affect the diagnosis of recurrence. The diagnostic performance of NPC images based on PET-CT imaging was evaluated using 42 cross combinations of six feature selection methods and seven classifiers. The optimal combination of feature selection and machine learning methods (the cross-combination fisher score [FSCR] + random forest [RF], FSCR + k-nearest neighborhood [KNN], FSCR + support vector machines [SVM] with radial basis function kernel [RBF-SVM], and minimum redundancy maximum relevance [MRMR] + RBF-SVM) to identify the two diseases were obtained (AUCs of 0.883, 0.867, 0.892, and 0.883, respectively; sensitivity: 0.833, 0.864, 0.831, and 0.750, respectively; specificity 1, 1, 0.873, and 1, respectively). Compared with the standard uptake value (SUV), total lesion glycolysis, and other indices, radiomics showed a higher AUCs (0.867–0.892 vs. 0.817), although the difference was not statistically significant (p = 0.462–0.560) (Table 1).

Radiotherapy is the main treatment for NPC during the early stage and radiotherapy and chemotherapy are the primary treatments during in late stage (6). For patients with intensity-modulated radiotherapy, weight loss, tumor regression, and other factors can result in large dose errors when applying the originally planned irradiation (30). In such cases, adaptive radiotherapy may be a better treatment option. Adaptive radiotherapy is usually administered to patients during radiation therapy, and the processes of imaging, sketching, and planning are repeated. The current radiotherapy system presents a significant economic burden for patients, and treatment is time-consuming and labor-intensive. Patients who need adaptive radiotherapy should be identified before treatment to improve treatment response. Yu et al. (30) used tumor marker features in MRI images acquired before treatment, and feature modeling (using enhanced T1 and T2 images provided AUCs of the enhanced T1, T2, and combined model verification groups of 0.852, 0.750, and 0.930, respectively), which may offer a basis for determining patient eligibility for adaptive radiotherapy and developing personalized treatments to reduce dose error. For the treatment response of NPC patients with advanced local progression to induction chemotherapy, Zhao et al. (31) developed a nomogram that combined multi-sequence MRI features before treatment with clinical parameters to predict the treatment effect in non-epidemic NPC areas. The model (training and validation group C-index: 0.952 vs. 0.863) had better predictive ability than the model developed using clinical parameters alone (training and validation group C-index: 0.708 vs. 0.549). In addition, in a study by Piao et al. (32), the AUC of the combined model was highest (AUC: 0.905) with the separate modeling of ClusterShade_angle135_offset 4 and Correlation_AllDirection_offshel_SD features based on enhanced magnetic resonance sequence imaging (AUC: 0.804 and 0.762, respectively). The combined model of the two features can help to determine the sensitivity and drug resistance in patients undergoing neoadjuvant chemotherapy, which is crucial for treatment scheme selection and treatment plan modification in patients with NPC.

Most imaging studies have focused on the prognosis of NPC. These studies (33–36) demonstrate the effectiveness of conventional MRI in evaluating progression-free survival (PFS), disease-free survival, and overall survival in patients with NPC, in combination with clinical information, such as lymph nodes, Epstein-Barr virus, and tumor stage, which can guide personalized treatment selection and improve the quality of care. Ouyang et al. (33) calculated and analyzed the Radscore and found it can predict PFS as a biomarker. Shen et al. (34) developed five models based on different combinations of data: model 1: clinical data; model 2: overall staging; model 3: radiomics; model 4: radiomics + overall staging; and model 5: radiomics + overall staging + EB virus). Model 5 had a high C-index for predicting PFS (training group 0.805, validation group 0.874). Yang et al. (36) suggested a nomogram integrating lymph node, Dose Volume Histogram signature, reflecting planning score and TNM stage, that had a C-index of 0.811 for the prediction of PFS, which showed better performance than using TNM alone (C-index: 0.613). Furthermore, another study (37) used different combinations of PET, CT, and relevant clinical data to develop models and found that the model combining all three factors had the highest prediction performance (C-indices of the training and validation cohorts were 0.71–0.76 and 0.67–0.73, respectively). Another found that subregional radiomics analysis of NPC outperformed the whole tumor (C-index, 0.69 vs. 0.58) and the traditional AJCC (American Joint Committee on Cancer) staging system for PFS prediction (38).

Radiomics can predict not only treatment effects but also recurrence before treatment, which can improve treatment decision-making. For the prediction of recurrence, most studies use MRI images. Zhang et al. (39) developed models based on MRI radiomics to predict distant metastasis (AUCs of the training and validation groups: 0.827 and 0.792, respectively) and divided patients into low- and high-risk groups based on a risk cutoff score of 0.37 to indicate the risk of metastasis and determine the treatment strategy. A subsequent study (40) introduced a nomogram to radiomics to study local recurrence and found that the nomogram (C-index: 0.74) predicted recurrence more accurately than did radiomics and clinical variables (C-index: 0.59). The study by Raghavan et al. (41) preferred the prediction model, which not only predicted recurrence but also emphasized whether recurrence would occur in the form of local or distant metastasis. The AUC, sensitivity, and specificity of the local recurrence model were 0.82, 0.73, and 0.74, respectively, whereas those of the model for predicting distant metastasis were 0.92, 0.79, and 0.84, respectively. In addition, another study (42) combined machine learning with features extracted from MRI and applied different feature selection and classifier methods to determine the optimal combination (random forest + random forest), which laid a foundation for future studies of local recurrence and distant metastasis prediction combining MRI features with relevant clinical information. Li et al. (43) used radiomics with machine learning to analyze the radiation resistance of local recurrence (artificial neural network: 0.812; KNN:0.775; SVM: 0.732) using existing MRI data, which provided quantitative and objective evaluations of patients with NPC without requiring additional radiation exposure. Furthermore, NPCs with in-field recurrences could be differentiated from NPCs (AUCs: 0.727–0.835).

Radiomics can also be applied to the prediction of radiotherapy reactions the following radiotherapy for NPC. In a study of patients with acute xerostomia after radiotherapy (44), parotid CT images and saliva volume were acquired before, during, and after treatment to develop a model to predict changes in saliva volume after early radiotherapy (accuracy: 0.9220, sensitivity: 100%). The difference between the statistical and real values can then be used to predict the degree of dry mouth by predicting the amount of saliva. The diagnosis of radiation-induced brain injury in NPC mainly depends on MRI; however, MRI has limited use for early diagnoses and can only be used to evaluate morphological changes in late radiation-induced brain injury in the temporal lobe. Radiomics can examine microscopic characteristics, which can be used as markers as a basis for the treatment of early brain injury. Zhang et al. (45) developed three models combining machine learning and MRI radiomics; the AUCs of the validation group were 0.830 (95% confidence interval [CI]: 0.823–0.837), 0.773 (95% CI: 0.763–0.782), and 0.716 (95% CI: 0.699–0.733), respectively, which offers promise for applying radiomics to the study of related complications.

Because there are numerous methods to extract radiomics features, obtaining robust features is vital for the generalizability of radiomics models. Liang et al. (46) used two different feature extraction tools to extract features from different NPC MRI sequences. Different extraction methods had varying effects on the features, which may impact model development. Thus, the selection of stable features of the disease is key. In a study on PET-CT radiomics characteristics under different contrast agents, Lu et al. (47) selected [¹⁸F]-FDG and [¹¹C] choline to examine segmentation and discretization and revealed that discretization has a greater impact on features than does segmentation, and features extracted from [¹¹C] choline are more stable than those extracted from the [¹⁸F]-FDG contrast agent. Yang et al. (48) evaluated the reproducibility of features extracted from PET-MRI and found that a voxel size of 0.5 × 0.5 × 1.0 mm³ in PET, T2, and diffusion-weighted imaging data and a larger bin size allow the acquisition of stable characteristics. Although these studies focused on the definition and mode of feature generation, Lv et al. (49) analyzed the robustness of feature matrix parameters and found that poor absolute-scale robustness retained good diagnostic performance (Table 1).