Deciphering Clinicoradiologic Phenotype for Thymidylate Synthase Expression Status in Patients with Advanced Lung Adenocarcinoma Using a Radiomics Approach

We aimed to identify predictive clinicoradiologic characteristics of thymidylate synthase (TS) expression status in advanced non-squamous non-small cell lung cancer patients. We reviewed clinicoradiologic features of 169 patients stratified into TS-negative (n = 84) and TS-positive (n = 85) groups, including quantitative CT radiomic features of both primary lung and metastatic lesions from initial CT and PET. Clinical factors including age and smoking history were significantly associated with TS as well as radiomic features. The predictive performance for dichotomizing TS expression status was slightly higher when imaging features of primary lung lesions were added compared to the model based solely on the clinical features, but without statistical significance (10-fold cross-validated AUC = 0.619 and 0.581, respectively; P = 0.425). The predictive performance of clinicoradiologic parameters slightly increased with primary lung lesions only compared to the inclusion of metastatic lesions, but without statistical significance (10-fold cross-validated AUC = 0.619 and 0.554, respectively; P = 0.203). Overall survival was prolonged in the TS-negative group compared to the TS-positive group (P = 0.001). TS-negativity is a potential prognostic biomarker, and our study presents that although CT radiomic features have potential for predicting TS expression status, clinical significance is uncertain. The addition of radiomic features to clinical factors did not show significant improvement in predicting TS-negativity.

a biomarker that may be helpful in determining a treatment regimen or predicting prognosis in NSCLC patients, ultimately for improvement in survival rate, is needed.
Thymidylate synthase (TS) is a key enzyme in Deoxyribonucleic acid (DNA) synthesis and is also the main target of antifolate drugs including pemetrexed. Pemetrexed, which is increasing its therapeutic scope from second-line therapy to first-line and maintenance therapy, is a multi-targeted antifolate drug that inhibits at least three enzymes including TS involved in DNA synthesis and folate metabolism, resulting in antitumor effects 5,6 .
Previous studies have reported that lower TS expression level is associated with better clinical outcomes for pemetrexed-based chemotherapy in NSCLC [7][8][9][10] . A recent phase III study showed survival differences between histologic types; pemetrexed/cisplatin was superior for survival compared with gemcitabine/cisplatin in nonsquamous NSCLC, although it was inferior in squamous cell histology 11 . This result can also be explained by the higher TS expression level in squamous cell carcinoma compared with other histologic types including adenocarcinoma 12 . Moreover, several studies have shown the association between low TS expression level and prolonged overall survival in NSCLC regardless of treatment regimen. Therefore, TS expression itself may be a potential prognostic factor [13][14][15] .
Regarding the evaluation of TS expression, different methods have been used including immunohistochemistry (IHC) to detect TS expression at the protein level and real-time reverse transcriptase polymerase chain reaction (RT-PCR) at the messenger ribonucleic acid (mRNA) level. Although IHC has been widely used and has shown better correlation with objective response rate prediction in patients with lung cancer receiving pemetrexed-based treatments, factors like tumor heterogeneity, subjective scoring system, or cutoff value are considered limitations, for which cautious data interpretation is required 16 . Also, the assessment of TS expression requires a tumor biopsy, which may not be possible for every primary tumor site. Therefore, a noninvasive imaging biomarker aiding in the prediction of TS expression level would be of great clinical importance.
Thus, we aimed to evaluate the clinicoradiologic features of NSCLC stratified by TS expression status and to correlate those with the prognosis in order to identify useful predictive imaging characteristics of TS expression status and to help develop improved treatment strategies.

Materials and Methods
Our institutional review board approved this retrospective study with a waiver of informed consent (SMC #2015-07-149). This study was performed in accordance with the principles of the Declaration of Helsinki for medical research involving human subjects.

Patients.
We identified 315 patients who had been histologically diagnosed with advanced non-squamous NSCLC (stage IIIB, IV, or recurrent disease after complete resection) at Samsung Medical Center (Seoul, Korea) from July 2011 to January 2014, who were originally included in a biomarker-stratified randomized phase II trial 17 . All patients had immunohistochemical analysis results for TS expression of biopsy specimens and chest computed tomography (CT) within 3 weeks. Twenty-six patients who had no follow-up CT to evaluate the prognosis were excluded. Twenty-nine patients were excluded for CT review-related factors: non-contrast CT only, images that could not be used for quantitative CT analysis due to the raw dicom data error, or no measurable lesion on CT by RECIST 1.1 18 . Considering the availability of 18 F-fluoro-2-deoxyglucose (FDG)-positron emission tomography (PET) imaging, 169 of 260 patients were included in the present analysis. Four of 260 patients who had histology other than adenocarcinoma; pleomorphic carcinoma or not otherwise specified, were excluded due to prognosis or image-related factors described above. Therefore, the histology of all 169 patients included was adenocarcinoma.
All patients were stratified into TS-negative or TS-positive groups. Patients for each group were randomly assigned to receive platinum doublet chemotherapy with or without pemetrexed: either intravenous pemetrexed (500 mg/m 2 ; Alimta, Eli Lilly, Indianapolis, IN, USA) plus cisplatin (70 mg/m 2 ; Cisplan, Dong-A ST, Seoul, Korea) intravenously on day 1 or gemcitabine (1000 mg/m 2 ; Gemzar, Eli Lilly, Indianapolis, IN, USA) on days 1 and 8 plus cisplatin (70 mg/m 2 ) intravenously on day 1. The administration was repeated every 3 weeks, and patients received a maximum of 6 cycles of chemotherapy until disease progression, unacceptable adverse event, or decision by the patient or physician. The response was assessed by CT scans every two cycles according to RECIST 1.1 18 . Imaging and Analysis. Imaging characteristics of each lesion were evaluated using chest CT and the PET component of PET/CT. Dedicated chest CT images were obtained with several multi-detector CTs including 8-, 16-, or 64-detector row CT scanners. All CT exams were performed with contrast. Chest CT data were interfaced directly to a picture archiving and communication system (PACS) (Path-Speed or Centricity 2.0; GE Healthcare, Mt. Prospect, IL, USA), which displayed all image data on two monitors (1536 × 2048 matrix, 8-bit viewable grayscale, 60-foot-lambert luminescence). The monitors were adapted to view both mediastinal (width, 400 HU; level, 20 HU) and lung (width, 1500 HU; level, −700 HU) window images.
Three radiologists (H.Y.L., H.S.H., and E.Y.K.) blinded to the information about the TS group and treatment allocation performed independent reviews of the radiological data from all randomized patients with at least a baseline and one follow-up scan and evaluated treatment response prospectively in consensus. A maximum of five lesions including primary lung lesion for each patient were included and measured at baseline CT according to RECIST guidelines 18 .
A thoracic radiologist (S.W.L, with six years of experience in thoracic CT interpretation) who was unaware of the clinical data and histologic diagnoses retrospectively evaluated the initial CT scans. The size and opacity of the primary lung lesion, the presence/absence of pleural metastases and lymphangitic carcinomatosis, and the presence and location of distant metastases were evaluated as CT features. The opacity of primary lung lesion was characterized as either solid (lesions with increased opacity obscuring pulmonary vessels), non-solid (lesions with ground glass opacity preserving the bronchial and vascular margins), or part-solid (lesions with mixed solid and ground glass opacity components). For quantitative CT analyses, tumors were segmented by drawing a region of interest (ROI) covering the largest possible area of the whole lesion ( Figure S1). Not only the primary lung lesions, but also the metastatic lesions were segmented by drawing a ROI. We used Mricro software (http://www.cabi.gatech.edu/mricro/mricro/ version 1.40) to draw the ROIs. Next, voxel-based CT numbers were collected from lesion segmentations.
We computed radiomic features from raw imaging data over an ROI drawn by a thoracic radiologist (S.W.L, with six years of experience in thoracic CT interpretation) 19 . A total of 60 features were computed in our study(- Fig. 1). The features were classified into four categories: histogram-based, shape-based, gray level co-occurrence matrix (GLCM), and intensity size zone (ISZ) features. Histogram-based features were computed using the intensity (HU) distribution of a given ROI. These features reflect intensity information of a given ROI. For histogram features, we adopted 2 types of histograms as follows: (1) regular histogram of all intensity values within the ROI (19 features); (2) histogram based on inner and outer portions of the ROI (18 features). A given ROI was partitioned into inner ROI (2/3 of the whole ROI volume) and outer ROI (1/3 of the whole ROI volume) purely based on the volume. We hypothesized that partitioning the outer part of the ROI would better reflect the aggressiveness of the tumor than the total ROI, which may have an offset portion as a whole. As the volume of the tumor decreased with this partitioning procedure, some of the features were not analyzable separately for inner or outer portion of the tumor due to its small volume. Therefore, inner and outer ROIs each had nine features. The difference between inner and outer ROI value was defined as delta ROI, which may signify intra-tumor heterogeneity. A total of 9 features were included for both outer ROI and delta ROI. The total number of histogram-based features, including median, energy, min, max, and range, was 37 20 . Shape-based features reflect morphological data of the ROI. A total of 10 features including volume, surface area, convexity, and compactness were computed 19,21 . GLCM-related features consider intensity values of a neighborhood instead of one voxel. One could quantify how similar or dissimilar voxel intensities are within a neighborhood. Thus, GLCM-related features could be used to quantify texture data 22 . Image intensities were discretized to 256 level for robust computation of GLCM, which was consistent with the Freedman-Diaconis formula and other studies 23,24 . GLCMs were computed for 13 directions and the average of 13 matrices were used for feature computation. The GLCM had 11 features including auto-correlation, and the dissimilarity was computed 19 . ISZ features were also correlated to texture data, but could be quantified beyond the immediate neighborhood. They assumed that an ROI could be further divided into sub-regions with uniform intensity but variable size. ISZ could quantify how many sub-regions and how often certain sub-regions occur within the tumor 25,26 . Image intensities were discretized to 32 level for robust computation of ISZ matrix. We chose 32 bins as it provided sensitive matrix to quantify number of sub-regions 26 . A total of two features, intensity variability and size-zone variability, were computed. Details regarding mainly adopted features are provided in the Supplementary Table. To increase the reproducibility of results, most of our radiomic features were computed using open-source platform called PyRadiomics 27 , available at www.radiomics.io. Only for 11 features including percentile (2.5%, 25%, 50%, 75% and 97.5%), uniformity of positive gray level pixel values (UPP), density, mass, convexity, size-zone variance and intensive variance, we used in-house MATLAB code (Mathworks Inc., MA, USA). The source codes used for the analysis are provided in the supplementary data. PET/ CT was available in 169 patients. Details of PET/CT acquisition are described in Appendix S1.
For semi-quantitative analysis of FDG uptake from PET/CT, ROIs were placed over the most intense area of FDG accumulation. When nodular FDG uptake could not be assessed on PET component images of PET/CT, an ROI was drawn in a presumed nodular location based on CT component images of PET/CT. FDG uptake within the ROIs was calculated as the maximum standardized uptake value (SUVmax).
Thymidylate Synthase Expression Analyses. All specimens obtained from patients with advanced non-squamous NSCLC were analyzed for TS protein expression by IHC as described previously 10 . The specimens used in this study were 4-µm-thick sections of paraffin-embedded tissue, and monoclonal anti-TS (4H4B1, Invitrogen, Carlsbad, CA) was used for the staining of TS protein. IHC staining of TS was assessed by two investigators (S.Y.H. and J.H.). Based on a previous study using a modified H-score system by Sun et al. 10 , we stratified TS-positive or TS-negative tumors by the cutoff value of 10%. TS-positivity was defined as TS expression in more than 10% of tumor cells, and TS-negativity as TS expression in 10% or fewer tumor cells.

Statistical Analysis.
To evaluate the association between clinicoradiologic characteristics and TS expression status from primary lung lesions and mean value of both primary and metastatic lesions, logistic regression analysis was performed. Characteristics with P-value less than 0.05 were considered to be statistically significant.
For the analysis of progression-free survival (PFS) and overall survival (OS) according to TS expression status, we used log-rank test with Kaplan-Meier curves. Statistical analyses were performed using SPSS software (SPSS for Windows, version 11. Internal validation. To evaluate the predictive performance of the prediction model, 10-fold cross-validation (CV) procedure was used as follows: Step 1. The total data were randomly divided into ten equally sized subsets.
Step 2. A single subset was used as the validation data, and the remaining nine subsets were used as training data.
Step 3. The variables were selected using variable selection procedure based on random forests proposed by Genuer et al. in the training set 28 .
Step 4. Random forest was applied to the selected variables to fit a prediction model.
Step 5. A fitted prediction model was applied to the validation data and the predicted probabilities were calculated.
Step 7. After the cross-validation is complete, the predicted probability values of all patients calculated by 10 fold CV were combined together. A single ROC curve was drawn as in Simon et al.

Analysis Dealing With Only Primary Lung Lesions.
The association between imaging parameters and TS expression status was evaluated only for primary lung lesions. Considering the availability of PET/CT imaging, a total of 169 patients with primary lung lesions were included. Younger age and never-smoker status were significantly associated with TS negativity (P = 0.003 and P = 0.024, respectively). Higher skewness (P = 0.041) and lower kurtosis (P = 0.005) of the whole tumor and higher skewness outer (P = 0.020) and lower kurtosis outer (P = 0.016) of histogram-based features were also associated with TS negativity. The root mean square (RMS), mean outer , median outer , and maximum delta values from histogram-based features and dissimilarity and entropy of GLCM-based features were significantly associated with TS expression status ( Table 2)

Analysis of Combined Data of Primary Lung Lesions and Metastatic Lesions. The association
between the imaging parameters and TS expression status was evaluated for the primary lung lesions as well as the metastatic lesions, for which the mean value of primary and metastatic lesions from each patient was used for analysis. Eighty-four patients had more than one metastatic lesion other than the primary lung lesion. The mean number of total lesions in each of all 169 patients was 1.75. Younger age and never-smoker were significantly associated with TS negativity (P = 0.043 and P = 0.005, respectively). A lower CT attenuation value at the 25 th percentile and higher CT attenuation value at the 97.5 th percentile from histogram-based features were significantly associated with TS negativity (both P < 0.001). Inter-quartile range (IQR), RMS, mean outer , median outer , and maximum outer values from histogram-based features were also associated with TS expression status (Table 3). SUVmax from PET/CT was not significantly associated with TS expression status (P = 0.265).

Predictive Performance of Clinical and Imaging Parameters for Dichotomizing TS Expression
Status. To compare the predictive performance of clinical and imaging parameters for dichotomizing TS expression status, 10-fold cross-validation (CV) procedure was used. 10-fold cross-validated AUC of the prediction models using only clinical variables, both clinical variables and radiomic features from primary lung lesions, and both clinical variables and radiomic features from the combined data of primary and metastatic lesions were 0.581, 0.619, and 0.554, respectively (Fig. 2).
The performance of predicting the TS expression status was slightly higher when radiomic features of primary lung lesions were added compared to the model based solely on the clinical features, but without statistical significance (P = 0.425). The predictive model with primary lung lesions showed slightly higher AUC compared to the added consideration of metastatic lesions, but without statistical significance (P = 0.203).

Survival Analyses. The correlations between TS expression status and PFS and OS in 260 patients with
non-squamous NSCLC are shown in Fig. 3. The PFS was not significantly associated with TS expression status (P = 0.144). However, the OS was significantly prolonged in the TS-negative group compared with the TS-positive group (P = 0.001).  Table 3. The result of association between clinicoradiologic variables and TS expression status from the mean value of both primary and metastatic lesions. Abbreviations: IQR, interquartile range; RMS, root mean square. Unless otherwise indicated, data are median (interquartile range). * Data indicate the number of individuals (percentage).

Discussion
Recent studies have reported that low TS expression level is associated with better clinical outcome and longer overall survival [8][9][10][13][14][15]17 . Consistent with these previous reports, our study also suggests that TS-negativity is a good prognostic factor. A few studies have shown controversial results that high TS expression level was associated with improved survival 31,32 . These discrepancies are mainly due to tumor heterogeneity. In other words, the result from a small specimen may not be indicative of the TS status of the entire tumor. Additionally, the invasive process of obtaining a specimen from the tumor and the non-standardized procedures to evaluate TS expression level remain critical points. The development of antibodies to detect TS in a sensitive and quantitative assay made it possible to measure the TS expression level in tumor samples using IHC 33 . However, TS evaluation with IHC is a semi-quantitative and empirical method that depends on the detailed procedures and is therefore difficult to standardize. Considering these limitations in detecting TS expression level from tumor samples, predicting TS expression level based on clinicoradiologic factors determined in non-invasive studies such as CT or PET/CT images of the NSCLC patients is of great clinical importance. A previous study has described the clinicopathologic characteristics of TS-negative lung cancer, demonstrating that female, younger age, never-smoker status, and adenocarcinoma were more frequent in TS-negative patients 10 . Since these clinicopathologic characteristics are not specific for TS expression status, adding radiologic features may increase the predictive probability. Radiomic features from CT or PET are widely being explored by recent studies for their correlation with prognosis, treatment response or tumor phenotype including EGFR-mutation status in NSCLC patients [34][35][36] . Our study also presents that some CT radiomic features were univariately predictive of TS expression status. However, the addition of radiomic features to clinical factors did not show significant improvement in predicting TS-negativity. CT radiomic features have potential for predicting TS expression status, but there is no clinical significance beyond the data variability of clinicoradiologic features. SUVmax from PET/CT was not significantly correlated with TS expression status. To our knowledge, there has been no previous description of the radiologic characteristics of lung cancers with TS stratification, and our study is the first attempt to better understand both clinical and radiologic characteristics of TS-positive lung adenocarcinomas.
We determined that radiomic features including skewness, kurtosis (histogram-based features), entropy, and dissimilarity (GLCM-based features) from primary lung lesions reflecting tumoral heterogeneity were significantly correlated with TS expression level. Histogram-based features reflect intensity information of a given ROI and quantify intra-tumor heterogeneity. Lower skewness (a measure of asymmetry of a histogram) and higher kurtosis (a measure of flatness of a histogram) representing increased heterogeneity were observed in both the total and outer ROI values of TS-positive tumors. In NSCLC, intra-tumor heterogeneity reflected by CT texture analysis is known to be associated with MAPK tumor pathway and has potential as a prognostic indicator 37 . Win et al. suggested the intra-tumor genomic heterogeneity or tumor hypoxia as the probable reasons for the relationship between tumoral heterogeneity and prognosis 38 . In previous studies, both kurtosis and skewness were associated with v-Ki-ras2 Kirsten rat sarcoma viral oncogene homolog (KRAS) mutation, and epidermal growth factor receptor (EGFR) mutation status was also correlated with CT texture features 39,40 . Tumor heterogeneity reflects the distribution of CT pixel values of the tumor and is related to tumor aggression. Additionally, there is emerging evidence that tumor segmentation of NSCLC on CT may improve prediction of survival (e.g., necrosis in the core and proliferation along the boundary) 21 . In our study, we partitioned the outer 1/3 of the ROI, and the mean, median, skewness, and kurtosis of the outer ROI and the maximum of the delta ROI were significant predictors of TS expression status. We made an attempt to enhance the value of the tumor partitioning for more detailed evaluation of tumor texture analysis. For validation of these findings, tumor partitioning must be applied in a larger study.
Whereas histogram-based features reflect intra-tumor heterogeneity in terms of intensity, GLCM-based features, which are textural features, reflect the heterogeneity of spatial distribution of voxel intensities. These features consider intensity values of a neighborhood instead of a single voxel. In our study, entropy and dissimilarity (GLCM-based features) of primary lung lesions were significantly correlated with TS expression level. Entropy indicates the uncertainty of the GLCM, which reflects the randomness of the matrix. Dissimilarity indicates how different each element is in the matrix.
Recently, with the development of targeted therapies in NSCLC, determining tumor mutation status is an important part of diagnosis in many countries to improve treatment outcomes 41 . However, in advanced lung cancer patients, therapeutic decision-making based on the results from primary tumor specimens could be justified only in cases where the tumor mutation status is similar between primary and metastatic lesions 42 . On the contrary, it is not always possible to obtain primary tumor tissue due to the invasiveness of the procedure, and a metastatic lesion may be obtained instead of the primary tumor 43 . Nevertheless, there is controversy over the concordance between the primary and metastatic lesions of NSCLC with multiple mutation statuses, including EGFR, KRAS, p16, and p53 [44][45][46][47][48][49][50][51] . Generally, substantial mutation status concordance was observed between primary and metastatic lesions in previous studies, with the concordance varying from 60 to 100% 51 . Potential causes contributing to some level of mutation status discordance include the testing methodology, site of the metastatic sample, and tumor heterogeneity 51 . Tumor heterogeneity exists not only within a single tumor (intratumor), but also between tumors of the same type in a patient (intertumor) 42 . Although there is a close genetic relationship between the primary and metastatic lesions, cancer is an evolving systemic disease, and the primary and metastatic lesions could become divergent as they evolve. In a previous study by Shimizu et al. 52 , patients with EGFR mutations in both primary and metastatic tumors showed a higher disease control rate compared to the patients with EGFR mutations in primary tumors only (P = 0.062). This suggests that heterogeneity of mutation status between the primary and metastatic tumors may influence the treatment efficacy.
Allowing for these current concepts regarding primary and metastatic tumors, whether or not metastatic lesions should be included in the analysis of radiomic features is an important issue. At least, in NSCLC, there has been no attempt to investigate TS expression in metastases and compare it with the matched primary tumor. Several studies have been published on the correlation between TS expression in the primary tumor and metastatic lesions in colorectal cancer patients [53][54][55][56][57][58][59] , where the majority of the studies revealed a poor correlation of TS expression between the primary tumor and distant metastases [53][54][55][56]59 . Discordance of TS expression level in the primary tumor tissue with the outcome of FU-based chemotherapy for metastatic disease suggests that TS analysis must be performed on biopsy material obtained from the metastatic site or sites 60 . Here, our study demonstrated that clinicoradiologic characteristics of primary lung lesion did not significantly better reflect the TS expression status obtained from the primary lung lesion itself in comparison with the combined data of primary lung and metastatic lesions. However, the potential variability of TS expression between different tissues raises concern over the use of primary tumor tissue to predict treatment response or prognosis in metastatic disease. Further and preferably prospective studies are needed to evaluate TS expression in NSCLC metastatic lesions in comparison with matched primary lung lesions and to correlate findings with responsiveness or prognosis.
A potential limitation of our study is its retrospective design from a single institution. However, the patients originally participated in a phase II trial and were prospectively stratified into TS-positive or -negative groups. CT or PET/CT exams performed on a several different multi-detector CT scanners and various organs included for radiologic characteristics evaluation in metastases might be other potential limitations. Regardless of heterogeneity in CT scanners or protocols, we minimized the effect of CT acquisition to radiomic feature analysis by including CT images with thin slice thicknesses (2.5 mm or less) and standard reconstruction algorithm. Further studies with a larger sample size of patients from multiple institutions are required to validate our results.
In conclusion, TS-negativity is a potential prognostic biomarker, and our study presents that although CT radiomic features have potential for predicting TS expression status, clinical significance in uncertain. The addition of radiomic features to clinical factors did not show significant improvement in predicting TS-negativity.