Table 2.

MRI Radiomics in head and neck cancer

Author and year of publication	Patient sample size and disease stage	Time of imaging and study type	Therapy	Number of parameters derived	Method of feature grouping/selection and number of parameters derived	Outcome parameter	Result
Ouyang et al19 Aug 2017	N = 100 TC = 70; VC = 30 NPC Stage III-IVb	Pre-treatment retrospective MRI: T2 and T1c 970 features Median follow up time: 39.5 months	NM	Shape and size, First-order features, Texture, Wavelet 5: T1C GLCMcorrelation, GLCM_IMc T2: GLRLM,GLCM variance and GLCM homogeneity	Feature selection: LASSO (5) Rad score: was used to dichotomise patients into Low or high risk MVCRA to yield HR	Prognosis: PFS Compare clinical model with combined clinical + radiomics model	Radiomics a significant independent predictor of PFS PFS shorter in high risk Rad score patients.
Zhang et al20 Aug 2017	N = 113 TC = 80; VC = 33 NPC Stage III-IVb	Pretreatment Retrospective MRI: T2 and T1c 970 features	NM	Shape and size, First-order features, Texture, Wavelet (4 T1c and 4 T2 features)	Feature Selection: LASSO logistic regression (8) RAD score Data dichotomised: PFS 3 yrs- Yes or No	Predict progression: PFS using AUC Compare T1c, T2 sequences models individually with a radiomics model using both combined	Radiomic model using joint T1c and T2 yielded highest AUC TC and VC (compared to T1 c or T2 alone)
Zhang et al21 Aug 2017	N = 118 TC = 88; VC = 30 NPC Stage III-IVb	Pretreatment Retrospective MRI: T2 and T1c 970	NM	Shape and size, First-order features, Texture, Wavelet	Feature Selection: LASSO logistic regression Nomogram discrimination and calibration: Using C index	Prognosis: PFS Association b/w radiomics and clinical features using heatmaps	Radiomics significantly associated with PFS Radiomics plus clinical data: better in evaluating PFS than clinical data alone. Radiomic model using joint T1c and T2 better than T1c or T2 alone Radiomics plus TNM model outperformed TNM staging alone.
Wang et al22 Jan 2018	N = 120 (NPC stage II,III and IV)	Pretreatment Retrospective MRI: T1, T1c, T2w and T2wFS 591	2 cycles of IC every 3 weeks (Cisplatin, 5FU and Docetaxel)	Histogram, GLCM,GLRL, Gabor and wavelet features Data dichotomized: responder and non-responder to IC Internal validation.	Feature Selection: LASSO regression model five features from T1c; 15 features from combined model Association with response: Mann Whitney U test. ROC curves for discrimatory ability	Association b/w radiomics and response to IC Compared T1c with prediction Then compared model combining T1, T1c, T2w and T2wFS with prediction	T1c and combined sequences’ radiomics signature were independant predictors in discriminating response and non-response pretreatment. Combined model of all 4 MR sequences performed better than single T1c sequence.
Liu et al23 Dec 2015	N = 53 TC = 42; VC = 11 NPC	Pretreatment Retrospective 3T MRI: T1,T2 and DWI sequences. 126 features	RT with two cycles CCCT (Cisplatin)	GLCM GLGCM Gabor transform Intensity size zone matrix	Feature Selection: Fischer’s coefficient and PCA Supervised learning: two different algorithms used- kNN and ANN.	Evaluate T1,T2 and DWI combined with supervised machine learning algorithms in predicting tumour response to CRT	All three sequences showed predictive value. T1w texture parameters most accurate in differentiating responders vs non-responders
Jansen et al24 2016	N = 19 HNSCC DCEMRI scans at 1.5T	Pre- and intra treatment Retrospective DCE-MRI images, Ktrans and Ve.	CRT	Energy (E) and homogeneity	Forward sequential feature selection algorithm used, followed by logistic regression analysis, to determine the probability of prediction	Merits of texture analysis on parametric maps derived from pharmacokinetic modeling with DCE-MRI	Chemo-radiation treatment in HNSCC significantly reduces the heterogeneity of tumors. E of Ve was significantly higher in intra treatment scans, relative to pretreatment scans

ANN, Artificial neural network; CCCT, Concurrent Chemotherapy; CRT, Chemoradiotherapy; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging; HR, Hazard ratio; IC, Induction Chemotherapy; K^trans, volume transfer rate; MVCRA, Multivariable Cox regression analysis; NM, Not mentioned; NM, Not mentioned; RAD Score, Radiomics Score (Using linear combination of selected features weighed by relative coefficients); V^e, volume fraction of the extravascular extracellular space; kNN, kNearest neighbors.

Summary: Though on the outset 6 papers with sufficiently large sample sizes showing good performance of radiomics models may look encouraging, the fact that 4 of these appear to be same institution data with possibly overlapping patient cohorts warrant caution regarding the strength of evidence. Again, all papers were retrospective in design and evaluated the “predictive” role of radiomics as a biomarker, except the paper by Jansen et al²⁰ which was unique in that they compared pretreatment and intratreatment changes in texture analysis derived from DCE-MRI. Their finding of Energy of Ve increasing on treatment is interesting, however limited by the small sample size and lack of any internal or external validation of findings