Investigation of Radiation-Induced Toxicity in Head and Neck Cancer Patients through Radiomics and Machine Learning: A Systematic Review

Table 2

Summary of the main results of the selected studies. The table reports mean and standard deviation values of RQS which were attributed by authors to the included studies, along with a synthesis of the main results from each study (toxicity outcome’s prediction according to the performed radiomics and machine-learning analyses).

Study	Outcome	Imaging modality	Radiomic features	OAR	Patients number	Results	RQS (mean ± standard deviation)	RQS (mean, percentage)
Leng et al. [11]	Radiation brain injury	Diffusion tensor imaging (DTI)-MR	Fractional anisotropy map (one of the most common DTI parameters)	Brain (white matter)	77	Machine learning in DTI-MR can aid the early recognition of white matter injury	8 ± 4	22.2
Scalco et al. [12]	Parotid shrinkage	CT	7 texture/fractal features (mean, variance, entropy, homogeneity, entropy S2, fractal dimension, and volume cc)	Parotid glands	21	A significant decrease in mean intensity (1.7 HU and 3.8 HU after the second and last weeks, respectively) and fractal dimension (0.016 and 0.021) was found. Discriminant analysis, based on volume and fractal dimension, predicted the final parotid shrinkage (accuracy of 71.4%)	-1 ± 2	2.8
van Dijrk et al. [13]	Late xerostomia (at 12 months after RT)	Pretreatment T1w-MR	21 intensity and 43 texture features	Parotid glands	Total 93 (68 + 25, from 2 centres)	90th intensity percentile values (that is, high fat concentrations) associated with higher risk of xerostomia	18 ± 2	50
Abdollahi et al. [14]	Sensorineural hearing loss (SNHL)	CT	490 extracted features	Cochlea	47	10 features are associated with SNHL (AUC 0.88)	10 ± 5	27.8
Thor et al. [15]	Trismus at 1 one-year post-RT	Posttreatment T1 postcontrast MR	24 features	Masticatory muscles	20	Identification of mean dose/texture features candidate for trismus prediction	0 ± 1	0
van Dijrk et al. [16]	Late xerostomia (at 12 months after RT)	Pretreatment simulation FDG PET-CT	24 intensity and 66 texture features	Parotid glands	161	90th highest SUV values (high metabolic activity of the parotid gland) was associated with a lower risk of developing late xerostomia (xer12 m)	10 ± 1	27.8
Pota et al. [17]	Late xerostomia (at 12 months after RT)	CT	# Features	Parotid glands	37 (only 19 for xerostomia assessment)	Only preliminary data regarding the prediction of late toxicity, largely limited by the low sample size (n = 19)	4 ± 5	11.1
Gabrys et al. [18]	Late xerostomia (at 6–15 months and long-term toxicity outcome at 15–24 months after RT)	CT	# Radiomics and dosiomics features. Radiomic set: parotid shape (volume, sphericity, and eccentricity)	Parotid glands	153	Late xerostomia correlated with the contralateral dose gradient in the anterior-posterior (AUC = 0.72) and the right-left (AUC = 0.68) direction, whereas long-term xerostomia was associated with parotid volumes (AUCs >0.85), dose gradients in the right-left (AUCs >0.78), and the anterior-posterior (AUCs >0.72) direction	9 ± 1	25