The role of radiomics in tongue cancer: A new tool for prognosis prediction

Radiomics represents an emerging field of precision‐medicine. Its application in head and neck is still at the beginning.


Funding information
Italian Ministry of Health with Progetto di Eccellenza, Ricerca Corrente and 5x1000 funds variables were included in multivariable analysis and related to survival endpoints. Predictive models were elaborated (clinical, radiomic, clinicalradiomic models) and compared using C-index.
Results: In almost all clinical-radiomic models radiomic-score maintained statistical significance. In all cases C-index was higher in clinical-radiomic models than in clinical ones. ADC provided the best fit to the models (C-index 0.98, 0.86, 0.84 in loco-regional recurrence, cause-specific mortality, overall survival, respectively).
Conclusion: MRI-based radiomics in OTSCC represents a promising noninvasive method of precision medicine, improving prognosis prediction before surgery.

K E Y W O R D S
mobile tongue cancer, oral cavity tumor, precision medicine, prognosis prediction, radiomics

| INTRODUCTION
The accurate prognosis prediction represents the first objective for tailored therapy selection in oncology. In particular, there are different available oncological therapies (surgery, radiotherapy, adrotherapy, chemotherapy, immunotherapy) and the prognosis prediction aim to intensify the therapeutic regimen in selected cases, especially in planning adjuvant treatment. In recent years, the scientific community has focused on finding accurate, repeatable, and noninvasive biomarkers to better stratify patients' prognosis, improve the effectiveness of treatments and reduce related side effects (therapeutic index). Moreover, the recent growing volume of data and the complexity of decision-making processes has contributed to the birth of the "precision medicine," that considers the variability of patients and diseases. An emerging field of precision medicine is represented by radiomics, 1 which consists in a high throughput extraction of quantitative features from conventional radiological imaging, with the aim of quantitatively describe the cancer "phenotype." 2,3 Radiomics permits to analyze tumor heterogeneity in a noninvasive, economical, and repeatable way, offering a potential tool for precision medicine in cancer treatment. The quantitative data extracted from radiomics are called "features" and can be referable both to the macroscopic disease and to the tumor microenvironment, providing distinct and complementary information to the clinical models. Thanks to its potential implications, in recent years radiomics has generated growing interest from the scientific community. A recent review reported that the annual growth rate in the number of radiomics articles published between 2013 and 2018 was 177.82%. 4 Most of the included works are mainly focused on breast, lung, and prostate cancer. Conversely, in head and neck cancer its development and application are still at the beginning, and available studies are limited and mostly focused on nasopharynx and oropharynx cancers. 5 Furthermore, the great prevalence of radiomics analysis is based on computed tomography, while application on magnetic resonance imaging (MRI) is more complex and less studied in head and neck disease. 6 Despite the progress of the last few decades in diagnosis and treatment of oral tongue squamous cell carcinoma (OTSCC), the survival rates have improved in a marginal way. Nowadays the gold standard for OTSCC locoregional staging is the MRI, 7,8 while the gold standard of treatment is represented by surgery. 9 In case of intermediate-advanced pathological stage (pT≥3, pN+) and/or presence of other negative prognostic factors (e.g. perineural infiltration) adjuvant radiotherapy is recommended. 9 In case of lymphnode metastases with extranodal extension (ENE), also adjuvant chemotherapy should be performed. 9 Immunotherapy is generally reserved in case of relapse (loco-regional or distant) refractory to platinum-based regimens and with tumor expression of PD-L1 combined positive score ≥1. 10 Even with combined treatments, loco-regional relapses are frequent (40%) and the 5-year survival rate is still unsatisfactory (50%-60%). [11][12][13][14][15][16] The presence of metastases with extra nodal extension (ENE), positive surgical margins and higher depth of invasion (DOI) are considered as the most significant negative prognostic factors. 13,15,[17][18][19][20] Another negative factor recently considered in literature is the quantitative lymphnode burden: Zumsteg et al. 21 showed that even in patients with ENE, positive margins or both, only those with metastases in more than six lymphnodes had improved survival with postoperative chemoradiation. In general, the research of specifical prognostic factors aims to predict unfavorable outcomes in order to intensify therapeutic regimen in selected cases. In this regard radiomics could represent an added value in planning precision medicine.
In this study, we applied radiomics to preoperative MRI of patients with OTSCC, with the aim of defining potential prognostic biomarkers. We evaluated the association between radiomic features and oncological outcomes to preoperatively estimate the prognosis of these patients.

| MATERIALS AND METHODS
We performed a retrospective radiomic analysis on preoperative MRI of consecutive patients with OTSCC surgically treated at Istituto Europeo di Oncologia (IEO), Milano (Italy), between January 2010 and December 2019. The project was endorsed by Radiomic Board and Ethical Committee (ID-TRIAL 2520). All patients signed an informed consent for the anonymized use of data for clinical and scientific purposes.
• Neoplasia of the mobile tongue. We collected the following data for each patient: • For clinical variables, we performed a univariate analysis with cox regression models maintaining the oncological outcomes as endpoints (overall survival OS, locoregional recurrence LRR, cause specific mortality CSM). Patients' characteristics were summarized with descriptive statistics (frequency and percentages for categorical variables; median and Interquartile (IQR) range for continuous variables). Survival was calculated from the time of surgery to the date of death, recurrence, or last follow-up, whichever occurred first. Local recurrence includes relapses on T and/or N (distant metastases or second tumors were not included). Local recurrence was recorded as the time elapsed between the end of the treatment and the first appearance of disease.
All variables found to be statistically significant (p < 0.10) were included in the multivariable analysis (cox proportional hazard regression models) to assess their association as independent factors on survival endpoints. Risk estimates were quantified by hazard ratio (HR) and 95% confidence intervals (CI). Radiomic information was aggregated in the statistical analysis in terms of "radiomic score." Feature selection was carried out by first removing highly correlated features (Spearman ρ > 0.95) and zero-variance features. The remaining features were grouped in clusters (iterative hierarchical clustering algorithm) which grouped features with Spearman correlation ρ ≥ 0.75. From each cluster, the algorithm selected the feature mostly associated with each outcome (lowest univariate cox proportional hazard regression model p-value), removing the others. The task was iterated until no Spearman correlation ρ ≥ 0.75 among features was found, to guarantee the redundant features exclusion from the following steps. Then a model-based approach was adopted, and standardized features were selected by a multivariable cox regression LASSO model performing a further feature selection. The parameter λ value was chosen using 5-folds cross validation as the parameter maximizing the C-index. After estimating the model, for each patient, the radiomic score was defined as the product between the regression coefficients and their respective feature values. We elaborated different predictive models for each survival endpoint: clinical model (pre-and post-treatment), radiomic model, clinical-radiomic model (pre-and post-treatment).
The models were compared using the C-index, a measure of goodness of fit for binary outcomes in a classification setting, which ranges from 0 (very poor predictive model) to 1 (hypothetical perfect predictive model). To assess the validity of the estimates, for each model, a Bootstrap analysis with 5000 repetitions was performed to determine a CI for the C-index estimate. A likelihood ratio test (LRT) was performed to assess whether the radiomic model introduction improved the clinical models performance. Lastly, a sensitivity analysis was performed to overcome possible overfitting concerns: for each clinical-radiomic model an in-sample 5-fold-crossvalidation was implemented.
Analyses were performed using R(4.1.1) for Windows.

| RESULTS
Based on inclusion criteria, we retrospectively selected 79 patients ( Figure 1); main descriptive results are reported in Table 1 Figure 2. Figure 3 reported curves related to radiomic score in CE-T1 and ADC (T2W and DWI relative figures in Supplementary Material Figure S1). For each MRI sequence, the five statistical models below were estimated. Similar results were found for all sequences and complete results are reported in Supplementary Material (Tables S2-S4). In multivariable analysis, ADC provided the best fit to the models and related results are described in the text below. Tables 2-4 report results for LRR, CSM, and OS in ADC and also CE-T1 sequences, because it is the most widespread sequence (used in all MRI protocol in other facilities) and so it could be useful for external validation and future radiomics applications.

| Radiomic models
For LRR, CSM and OS, the feature selection algorithm yielded respectively 165, 170 and 174 features (ρ < 0.75). The radiomic score obtained after LASSO-cox regression T A B L E 1 Main characteristics of the study population: preoperative, surgical, histopathological, and follow-up data.  In the post-treatment clinical-radiomic model, radiomic score (HR = 4.37, p < 0.001) and pN+ ECE+ versus c/pN0 (HR = 5.08, p = 0.04) were the only variables significantly associated with the endpoint. After radiomic score was added, post-treatment clinical radiomic model C-index significantly (p < 0.001) increased to 0.86 (95%CI [0.79, 0.91]). With T2W sequences, radiomic score was only borderline associated with CSM at multivariable analysis (HR = 7.86, p = 0.06, Supplementary Material Table S3) providing no significant increase in C-index after its inclusion in the model (p = 0.06).

| Post-treatment models
A significantly better OS was observed for male patients (HR = 0.31, p = 0.003), while age at diagnosis and pathological status of lymphnodes pN+ ECEÀ and pN+ ECE+ versus c/pN0 were associated with a worse  Table S4) providing no significant increase in C-index after its inclusion in the model (p = 0.09).

| DISCUSSION
In this study, we investigate whether radiomics could represent an added value to clinical variables in terms of prognosis (even predictive) in OTSCC. The protocol for proper MRI in this setting of patients includes T1-and T2-weighted sequences, CE-T1 with fat saturation and optionally DWI and ADC. The main advantage of DWI and ADC in the staging of OTSCC is the ability to distinguish between tumor and peritumoral inflammation. 24 We included all these sequences to access whether one sequence could fit better with radiomics and for all models the integration of radiomics with clinical variables improved the patients' outcome prediction. The highest value of C-index was found in ADC for LRR (0.98). Unfortunately, DWI sequences and ADC maps are optional and not performed in all radiological centers, limiting its potential application. There is a need to introduce DWI sequence and ADC map in standard protocol of MRI in OTSCC for its recognized added value in radiological staging. 24 Basing on our results, it could be also useful for prognosis prediction, opening the horizon for other radiomic studies. However, it must be taken into account that features can be influenced by the variety of scanning protocols and MRI-equipment, therefore MRI acquisition protocols needs to be standardized to try to validate radiomics models across different Facilities. Anyhow, our data showed excellent results also in CE-T1 sequences, which are commonly used, being easily available and interpretable. This allow radiomics to be applied to wider number of patients, regardless of the center where the MRI is performed.
The radiomic model related to data acquired in the preoperative phase. This information was retrieved before the surgery, unlike most other prognostic factors which are accessible only after pathological examination. The "pre-treatment information" represents a key point of precision medicine and tailored treatments.
In the pre-treatment clinical models, for OS, multivariate analysis confirmed that age, female sex and Stages III-IV, represent the main negative prognostic factors; adding the radiomic score to the model, the C-index was F I G U R E 3 Kaplan-Meier plots for ADC (upper panels) and CE-T1 (lower panels) sequences according to high and low radiomic score, dichotomized based on its median value. Analogous curves relating to T2W and DWI are reported in Supplementary Material Figure S1.
[Color figure can be viewed at wileyonlinelibrary.com] T A B L E 2 Loco-regional recurrence regression   improved, and the stage was no more significant. For LRR pN+ ECE+ versus c/pN0 resulted to be the only significant variable associated with a higher risk of recurrence for OTSCC, as already reported in previous studies. 13,17,20,25 For CSM only Stage III-IV versus I-II was significantly associated with a worse outcome. After adding the radiomic score in LRR and CSM in multivariate analysis only the radiomic score was significant with a Cindex improvement. The lack of significance of stage and pN+ ECE+ at multivariable analysis including radiomic score suggests that radiomic features may incorporate information related to stage and to advanced disease, thus becoming the main outcomes predictor. For ADC, Stages III-IV have a significantly higher radiomic score, and pN+ ECE-have a significantly lower radiomic score value with analogous results for CE-T1, DWI and T2W sequences. The combination of clinical variables routinely collected with radiomic features quantitatively extracted from pretreatment MRI, can increase predictive accuracy for recurrence and mortality prediction. Radiomics models result independent factors in prognosis prediction, regardless of standard clinical factors and treatment.
To the best of our knowledge, this is the first study of radiomics for OTSCC that includes multiple MRI sequences and survival outcomes as clinical endpoints. At the current time only four studies focused on MRIbased radiomics in oral cavity cancer and only one focused on survival outcomes. 26 Frood et al. 27 (2018) conducted a retrospective study on 115 cases of oral cavity SCC to detect MRI radiomic textures indicative of lymphadenopathy with ENE. Nodal entropy derived from CE-T1 was significant in predicting ENE and nodal entropy combined with irregular boundary was the best predictor of ENE.
Ren et al. 28 (2020) applied radiomics on ADC MRI map of 88 patients with OTSCC distinguishing the degree of differentiation. Also Yu et al. 29 (2021) conduced an analysis on 127 OTSCC determining the degree of differentiation through radiomics on MRI fat-suppressed T2W images. Wang et al. 26 (2021) conducted a retrospective study on 236 patients with tongue cancer. MRI based radiomics of primary tumor and peritumoral tissue resulted an independent indicator for poor disease-free survival and overall survival.
A review published by Jethanandani et al. 6 in 2018 stated that MRI-based radiomics in head and neck district is currently a "young" and unexplored field for several reasons: anatomical complexity of head and neck region, lack of a common ontology, insufficient data source, impact of different scanners, acquisition parameters and image analysis methodologies are potential sources of bias. Furthermore, the multiparametric MRI analysis introduces the challenges related to each sequence, which can exhibit different properties (spatial resolution, signal/noise ratio, possible signal anisotropy). 30 We are aware of the limitations of our study, mainly including suboptimal subject numbers, lack of internal and external validation. In this study, we adopted methodological choices aimed at maximizing the robustness of the extracted features. We tried to overcome the problem of image variability through rigorous MRI selection: all exams were acquired on the same scanner with fixed acquisition parameters (TE, TR, voxel size). We calculated radiomic features from different image types (wavelet, Laplacian of Gaussian, squared, square root, logarithm, exponential, gradient, local binary pattern maps), identifying the procedure that maximizes the ability to predict clinical endpoints. We performed insample cross validation to validate our statistical models. Furthermore, the relatively small sample size avoided performing an external validation of our models, rising a possible overfitting. We performed several analyses to address this point, including 5-cross-validation of the parameter λ, bootstrap resampling to obtain the 95%CI and finally an in-sample 5-fold-cross-validation with calculation of average C-index on the test sets, providing similar results (Supplementary Table S5). Our analysis underlined that MRI-based radiomics in clinical practice provides important pre-treatment information in OTSCC contributing to the creation of highly individualized algorithms for treatment and follow-up. Future developments of this project will be founded validation in a multicentric study.

| CONCLUSIONS
In conclusion, radiomics application to preoperative MRI of patients with OTSCC represents a promising and noninvasive method of precision medicine. The integration of radiomics to standard clinical models has the potential to increase the accuracy in estimating the prognosis of these patients, providing accurate information before surgery. Furthermore, our results show that radiomics models are independent factors in prognosis prediction, regardless of standard clinical factors and type of surgery treatment. The incorporation of radiomics into clinical practice will likely contribute to improving precision medicine in patients with OTSCC in the future. To achieve these objectives, external validation of our data and other prospective studies of adequate numbers will be needed in the future.

AUTHOR CONTRIBUTIONS
All persons designated as authors participated sufficiently in the work to take public responsibility for the content of the manuscript: conception, design, analysis and interpretation of data, draft of the manuscript, revision for important intellectual content and final approval of the version to be published.

CONFLICT OF INTEREST STATEMENT
The authors declare no conflicts of interest. The institution of some authors (European Institute of Oncology) was partially supported by the Italian Ministry of Health with Progetto di Eccellenza, Ricerca Corrente and 5x1000 funds. Stefania Volpe MD and Lars Johannes Isaksson MSc are PhD students within the European School of Molecular Medicine (SEMM), Milan.