The radiomics features could be used to predict the distinct metastases in nasopharyngeal carcinoma patients

Background For NPC patients, distant metastasis is now the main reason for treatment failure. The patients with distinct metastases need different therapeutic regimen and have distinct prognosis. Radiomics might help us predict the type of metastases effectively. Methods The MRI data of seventy non-metastatic NPC patients who develop distant metastasis within ve years after treatment were collected and 4410 radiomics features for each patient were extracted by PyRadiomics. Every radiomics feature were compared among patients with distinct metastases and tested by the receiver operating characteristic curve. Results Twenty features have signicant differences between the bone metastases cohort and lung metastases cohort, two features have signicant differences between the bone metastases cohort and liver metastases cohort, one feature has signicant differences between the bone metastases cohort and multiple metastases cohort and sixty-seven features have signicant differences between the lung metastases cohort and liver metastases cohort. Six T2WI features could identied lung metastases from bone metastases and liver metastases effectively (AUC=0.851~0.896), two T1WI features and one CE T1WI feature could respectively identied bone metastases from liver metastases and multiple metastases effectively (AUC=0.779~0.821). Conclusion The results indicate that the radiomics features could reect some of the characteristics of distinct metastases and have potential to be used as predictor of distinct metastases. The Radiomics features which could classify the metastases effectively.


Background
Nasopharyngeal carcinoma (NPC) has a unique geographical distribution, with the highest incidence in Southeast Asia.More than 70% of NPC patients are diagnosed as locoregionally advanced disease due to its hidden pathogenic site [1][2][3][4].Currently the standard treatment for advanced NPC patients is radiotherapy combined with chemotherapy.Locoregional control has been improved substantially due to the use of intensitymodulated radiotherapy and combined chemoradiotherapy [5].Moreover, approximately 30% to 40% of advanced NPC patients eventually develop distant metastasis after treatment and distant metastasis is now the main reason for treatment failure [6].For advanced NPC patients, the common metastases were bone metastases, lung metastases, liver metastases, mediastina metastases and multiple metastases [6].The patients with distinct metastases need different therapeutic regimen and have distinct prognosis.
The tumor-node-metastasis (TNM) staging system is the key determinant for prognostic prediction and risk strati cation for treatment decisions.Even though NPC patients with the same TNM stage receive similar treatments, their clinical outcomes vary greatly.Hence, the current anatomical-based staging system is not su cient to predict the site of metastases.If we could predict the site of the metastases and prevent before the occurrence of metastases, the patients would bene t from it and the survival time of the advanced NPC patients would be prolonged.
Recently, the radiomics emerged and it enabled image features to be extracted from standard-of-care medical imaging and applied within clinical-decision support systems and has signi cantly improved the diagnostic, prognostic, and predictive accuracy of various tumors [7][8][9][10][11][12][13][14][15][16][17].Therefore, we wanted to see if radiomics could help us predict the site of the metastases in patients with advanced nasopharyngeal carcinoma.This study was approved by the ethics committee of the A liated Foshan Hospital of Sun Yat-Sen University.All the participants provided written informed consent.The research protocol complied with the guidelines outlined in the Declaration of Helsinki.The 8th AJCC staging system was used as a standard manual for tumor staging and the related clinical characteristics were collected including age, gender, T stage, N stage, Epstein-Barr virus (EBV) DNA copy number before treatment, lactate dehydrogenase (LDH), VCA-IgA, EA-IgA, C-reactive protein, white blood cell (WBC), platelet (PLT), lymphocyte (LYM), monocyte (MONO), hepatitis B surface antigen (HBsAg) and Treatment regimen.

Follow-up and clinical endpoint
After completion of radiation and chemotherapy, all patients were followed up every 3 months in the rst 2 years, every 6 months from year 3 to year 5, and annually thereafter.Follow-up was performed with endoscopy, head and neck computed tomography, or MRI.The follow-up period was de ned as the time from the date of diagnosis to the date of the last follow-up or death.

MRI acquisition and Image assessment
All patients underwent MRI with a 1.5 T system (Signa CV/i; General Electric Healthcare, Chalfont St. Giles, United Kingdom) or 3.0 T system (Siemens Magnetom Tim Trio, Erlangen, Germany), employing the fast spin-echo (FSE) technique.Axial T1-weighted images (T1WI), contrast-enhanced T1-weighted images (CE T1WI), and T2-weighted images (T2WI) were retrieved from the picture archiving and communication system (PACS) and loaded in AnalyzePro (https://analyzedirect.com/analyzepro/) for semi-manual segmentation, without any preprocessing or normalization.The MRI scans were evaluated by two independent radiologists with over 10 years' experience in assessing MRI scans of patients with NPC.Differences were resolved by consensus.

Statistical analysis
The clinical characteristics of the cohorts were compared using the Mann-Whitney U test for continuous variables and Chi-square test for categorical variables and two-sided P values <0.05 were considered to be statistically signi cant.The receiver operating characteristic (ROC) curve was used to illustrate the diagnostic ability of a binary classi er system.The area under the curve (AUC) was calculated by the Hanley and McNeil method.These analyses were performed using Graphpad Prism software (version 5.1) and SPSS (version 20.0).

Result Clinical Characteristics
The clinical and pathological features are listed in Table 1.The median follow-up period was 40.5 months (range: 6.2-72.2months) for all patients.All patients nally developed distant metastasis during the 5-year follow-up, twelve of them developed bone metastases, sixteen of them developed lung metastases, thirteen of them developed liver metastases, three of them developed mediastina metastases and the other developed multiple metastases.Radiomics features have signi cant differences among the patients with distinct metastases.
We compared these radiomics features among patients with distinct metastases and found that twenty features have signi cant differences between the bone metastases cohort and lung metastases cohort (Figure 1B), two features have signi cant differences between the bone metastases cohort and liver metastases cohort (Figure 1C), one feature has signi cant differences between the bone metastases cohort and multiple metastases cohort (Figure 1D) and sixty-seven features have signi cant differences between the lung metastases cohort and liver metastases cohort (Figure 1A).Moreover, one feature called "T2_log-sigma-5-0-mm-3D_glcm_ClusterShade" was higher in bone metastases cohort than other metastases cohort signi cantly (Figure 1E).These results indicated that MRI features could re ect some of the characteristics of distinct metastases and they might classify distinct metastases effectively.

Radiomics features could be used to classify distinct metastases
In order to explore which radiomics feature could classify distinct metastases, we performed ROC curve analysis and found that radiomics features could be used to classify distinct metastases effectively.As shown in Table 2 and Figure 2, six T2WI features could identi ed lung metastases from bone metastases and liver metastases effectively (AUC=0.851~0.896),two T1WI features and one CE T1WI feature could respectively identi ed bone metastases from liver metastases and multiple metastases effectively (AUC=0.779~0.821).Moreover, one T2WI feature could identi ed bone metastases from other metastases (AUC=0.739).Then we created 1,000 different datasets by bootstrapping (80% resampling) and validated these selected features and we found that most of them were robust (Table 2).These results indicate that the radiomics features have potential to be used as predictor of distinct metastases and help to the interfere treatment before metastases in NPC patients.
The Radiomics features which could classify the metastases effectively.

Discussion
We undertook this study to explore the value of radiomics in predicting the site of distant metastasis in NPC patients.We collected the MRI data of seventy non-metastatic NPC patients who develop distant metastasis within ve years after treatment and 4410 radiomics features were extracted from each MRI patient by PyRadiomics.We compared the radiomics features among patients with distinct metastases and found that radiomics features have signi cant differences among these groups and some features could be used to predict the type of metastases effectively.These results indicate that the radiomics features have potential to be used as predictor of distinct metastases in NPC patients.
Currently, several studies have assessed the performance of radiomics signature in clinical staging, prognostic prediction, response to treatment, and discrimination of phenotype for several types of tumors [8][9][10][11][12][14][15][16].For Nasopharyngeal carcinoma patients, distant metastasis after treatment is one of the most pressing issues and predicting the site of distant metastasis could be bene t to scheme of treatment.In 2018, a study reported a gene expression-based signature to predict distant metastasis in locoregionally advanced nasopharyngeal carcinoma [5] and they develop a nomogram to predict an individual's risk of distant metastasis.In that study they focused on the risk of distant metastasis and didn't group the patients according to the site of metastasis.Our study indicated that radiomics signature might be a good choice for predicting the site of metastasis because it allowed noninvasive and quantitative characterization of the tumor.
At present, the treatment is similar before the metastasis is diagnosed but the patients with bone metastasis have better outcome than liver and lung metastases.If the site of distant metastasis could be predicted, we could provide aggressive treatment regimen for the patients who have high risk of liver or lung metastasis to reduce the risk.The treatment might be different before the metastasis is diagnosed in the future to improve the prognosis of patients.
We collected the MRI data of 778 NPC patients who were treated at Sun Yat-Sen University Cancer Center from January 2009 to December 2012.Seventy patients developed distant metastasis during the 5-year follow-up.Limited by the number of patients, we couldn't validated the selected features in independent cohorts.We didn't know whether the features have the similar e ciency in distinct hospitals.We are trying our best to build cooperation with other radiologists and looked forward to getting diverse data to calibrate the ndings in this study.

Conclusions
In this study we focus on the radiomics features of distinct metastases and found signi cant differences.
They even could be used to predict the type of metastases effectively.These results indicate that the radiomics features have potential to be used as predictor of distinct metastases in NPC patients.The radiomics features which have signi cant differences between (A) the lung metastases cohort and liver metastases cohort, (B) the bone metastases cohort and lung metastases cohort, (C) the bone metastases cohort and liver metastases cohort, (D) the bone metastases cohort and multiple metastases cohort, (E) bone metastases cohort and other metastases cohort.

Patients
This study included 70 non-metastatic NPC patients who were treated at the A liated Foshan Hospital of Sun Yat-Sen University from January 2009 to December 2012 and nally developed distant metastasis during the 5-year follow-up.The Inclusion & exclusion criteria were as follows: (a) pathological diagnosis of NPC; (b) without evidence of distant metastasis; (c) underwent radical IMRT according to the standard guidelines; (d) without a primary tumor in other parts of the body; (e) complete clinical and detailed MRI report data were available; (f) developed distant metastasis after treatment during the 5-year follow-up.

Table 1 .
The clinical characteristic of the NPC patients in this study.