Table 1.

Radiomic studies in HNC, studying tumour control and normal tissue toxicity.

Study Type/Goal	Imaging Modality/Radiomic Features	Observations/Conclusions
Radiomic Studies for Outcome Prediction: Tumour Control
Staging and risk stratification model in oropharyngeal carcinomas (Cheng et al., 2013) [23]	PET/CT imaging Grey-level co-occurrence matrix, uniformity, and coherence	A risk stratification strategy was developed based on total lesion glycolysis (TLG) and uniformity. TLG, uniformity, and HPV positivity are significantly associated with overall survival.
Prognostic imaging biomarkers for overall survival (Parmar et al., 2015) [24]	CT imaging First-order intensity statistics, shape, and volume textural featuresArea under receiver operating characteristics curve (AUC) used to quantify the prognostic performance of different feature selection	Three feature selection methods—minimum redundancy maximum relevance, mutual information feature selection, and conditional infomax feature extraction had high prognostic stability and performance for the prediction of overall survival.
Prediction of local control using pre-treatment PET vs. CT (Bogowicz et al., 2017) [6]	PET/CT imaging CT density, HLH intensity, grey-level size zone texture matrices, and spherical disproportion	The model overestimated tumour control probability in high-risk patients. Combined PET/CT added no extra value when compared to either imaging method alone.
Risk prediction models of locoregional recurrences and distant metastases (Vallieres et al., 2017) [7]	CT imaging Large zone high grey-level emphasis; zone size non-uniformity	No significant correlation was found between radiomic features and locoregional recurrence. Image-derived features combined with clinical variables offer the highest predictive values.
Risk prediction models of all-cause mortality, local failure & distant metastasis (Folkert et al., 2017) [8]	CT imaging Image features: statistical, shape, and texture	Multiparametric models had the strongest predictive power. The local failure model demonstrated robustness when conveyed onto independent patient cohorts.
Radiomic Studies for Outcome Prediction: Normal Tissue Toxicity
Assessment of structural changes in parotid glands (Scalco et al., 2013) [9]	CT imaging Textural features and gland volume	Variations in mean intensity and fractal dimension (after the second and last week of radiotherapy) were the best predictors of parotid shrinkage.
Early prediction of parotid shrinkage and toxicity (Pota et al., 2017) [10]	CT imaging Textural features, spatial patterns, fractal dimensions, and gland volume combined with fuzzy classification	The final parotid shrinkage rate strongly correlated with 12-month xerostomia: glands that presented strong volume variation post-radiotherapy could be less affected by late xerostomia.
Prediction of radiation-induced xerostomia and sticky saliva (Van Dijk et al., 2017) [11]	CT imaging Geometric features, CT intensity, and textural features	Prediction of 12-month xerostomia and sticky saliva were improved by the addition to the initial CT image biomarkers of the short-run emphasis (quantifies heterogeneity of parotid) and of the maximum CT intensity of the submandibular gland (gland density).
Prediction of late xerostomia using parotid gland fat (Van Dijk et al., 2018) [12]	MR imaging T1-weighted MR-image-based intensity (90th intensity percentile) and textural features	The ratio of fat-to-functional-parotid-tissue is associated with 12-month xerostomia. MR-based radiomics improved prediction.
Prediction of severe late xerostomia (Nardone et al., 2018) [25]	CT imaging Textural features: grey-level co-occurrence matrix (GLCM), neighbourhood grey-level dependence matrix (NGLDM), grey-level run length matrix (GLRLM), grey-level zone length matrix (GLZLM), sphericity, indices from the grey-level histogram, and parotid volume	Parameters with the strongest correlation with severe chronic xerostomia: V30, mean dose, kurtosis, grey-level co-occurrence matrix (GLCM), and run length non-uniformity (RLNU). CT texture analysis could allow for the enhancement of dose constraints to organs at risk to avoid severe side effects.
Multivariable modelling study of chemotherapy-induced hearing loss (Abdollahi et al., 2018) [21]	CT imaging Textural features with highest predictive power: intensity histogram (IH) and grey-level co-occurrence matrix (GLCM)	Ten machine learning classifiers used for radiomic feature selection, classification, and prediction, with over 70% accuracy. No single algorithm showed superiority for all problems. CT image features of cochlea can serve as biomarkers for predicting hearing loss after therapy.
Early prediction of acute xerostomia during therapy (Wu et al., 2018) [26]	CT imaging Histogram-based features: mean CT number (MCTN), volume, skewness, kurtosis, and entropy	Daily CT images analysis during IMRT indicated that changes in gland volume or MCTN are not correlated with grade of xerostomia if considered separately but when combined in a CT-based xerostomia score model. This can predict severity with a precision of 100% at the 5th week of therapy.
Assessment of whole-brain white matter injury after radiotherapy (Leng et al., 2019) [22]	MR diffusion tensor imaging Fractional anisotropy (FA) and FA skeleton matrix	Post-therapy decreased FA values (that quantify the degree of water diffusion in cerebral white matter) showed microstructure damage of the white matter. Radiation brain injury in HNC patients can be quantified using MR diffusion-tensor-based radiomics.
Predictive model of acute radiation-induced xerostomia (Sheikh et al., 2019) [13]	CT and MR imaging Shape, first-order statistics, grey-level co-occurrence matrix (GLCM), grey-level run length matrix (GLRLM), and grey-level size zone matrix (GLSZM)Model performance improved when DVH was combined with CT and MRI features	Higher-order texture features for salivary glands were key predictors of xerostomia. MRI: patients with xerostomia appear more heterogeneous and hypointense. CT: submandibular glands of patient with xerostomia appear more hypodense and heterogeneous. Baseline CT and MRI features can potentially reflect baseline salivary gland function and the risk of radiation-induced effects.
Longitudinal study on post-radiotherapy parotid gland changes in nasopharyngeal cancer patients (Wu et al., 2020) [14]	MR and ultrasound (US) imaging MRI features: volumeUS features: echogenicity and hemodynamic parameters (resistive index, pulsatility index, and peak diastolic and end-diastolic velocity)	Parotid and submandibular gland shrinkage associated with xerostomia was observed post-radiotherapy (IMRT), with most significant changes detected after 6 months. Mild correlation found between gland dose and post-radiotherapy gland volume.