Applications of radiomics in precision diagnosis, prognostication and treatment planning of head and neck squamous cell carcinomas

Table 1 Prediction of HPV status based on radiomics features of HNSCC tumors

Authors, year	Sample size, cancer type	Ground truth	Imaging modality	ML classifier	Metric: maximum performance ^a
Bogowicz et al. 2017 [50]	Train: 93, HNSCC Test: 56, HNSCC	p16	Contrast CT	Logistic regression	Test-AUC: 0.78
Buch et al. 2015 [48]	Total: 40, OPSCC	Not reported	Contrast CT	n/a ^b	n/a ^b
Fujita et al. 2016 [49]	Total: 46: non-OPSCC	Not reported	Contrast CT	n/a ^b	n/a ^b
Huang et al. 2019 [51]	Train: 113, HNSCC Test: 53, HNSCC	Train: HPV RNA ^c Test: p16	Contrast CT	LASSO-regularized logistic regression	Nested CV-AUC: 0.73 Test-AUC: 0.76
Leijenaar et al. 2018 [52]	Train: 628, OPSCC Test:150, OPSCC	p16	Contrast CT	LASSO-regularized logistic regression	Test-AUC: 0.70–0.80 ^d
Mungai et al. 2019 [53]	Total: 50, OPSCC	Not reported	Contrast CT	Logistic regression	n/a ^e
Parmar et al. 2015 [54]	Train: 136, OPSCC and LSCC Test:95, OPSCC	Not reported	Contrast CT	Logistic regression	Test-AUC: 0.60
Ranjbar et al. 2018 [55]	Total: 107, OPSCC	HPV DNA-ISH	Contrast CT	Diagonal quadratic discriminant analysis	LOOCV-AUC: 0.80
Yu et al. 2017 [56]	Train: 150, OPSCC Test:165, OPSCC	p16	Contrast CT	Logistic regression	CV-AUC: 0.75 test-AUC 1 ^f: 0.87 test-AUC 2 ^f: 0.92
Zhu et al. 2018 [57]	Total: 126, HNSCC	Not reported	Contrast CT	Random forest	CV-AUC: 0.71

^a The reported performance pertains to pure imaging feature-based HPV classification (i.e. models with clinical features were not considered)
^b A t-test was used to evaluate differences in texture parameters between HPV-positive and HPV-negative cases
^c The VirusSeq-software was used to detect strain-specific HPV RNA sequences in whole-transcriptome sequencing data [51, 58]
^d This study evaluated the impact of CT artifacts on the HPV classification performance. A test set AUC performance of 0.8 was achieved after exclusion of all artifact-affected cases from both the training- and test set. The test AUC ranged between 0.70 and 0.80 for all evaluated dataset combinations, including those with artifacts, and was not significantly different for all tested models
^e The logistic regression model was trained and tested on the same dataset without feature selection or cross validation, which is prone to overfitting, and overestimation of classification accuracy
^f Study reports results of winning submission of radiomics competition, wherein 165 test cases were split into two test sets
AUC Area under the receiver operating characteristics curve, CV Cross validation (of total set or training dataset), DNA-ISH DNA in situ hybridization, HNSCC Head and neck SCC, LOOCV Leave one out cross validation of total set, LSCC Laryngeal SCC, OPSCC Oropharyngeal SCC, Test Independent test dataset, Total Only one dataset used, Train Training dataset

ISSN: 2059-7347