Pre-Operative Prediction of Mediastinal Node Metastasis Using Radiomics Model Based on 18F-FDG PET/CT of the Primary Tumor in Non-Small Cell Lung Cancer Patients

Purpose: We investigated whether a fluorine-18-fluorodeoxy glucose positron emission tomography/computed tomography (18F-FDG PET/CT)-based radiomics model (RM) could predict the pathological mediastinal lymph node staging (pN staging) in patients with non-small cell lung cancer (NSCLC) undergoing surgery. Methods: A total of 716 patients with a clinicopathological diagnosis of NSCLC were included in this retrospective study. The prediction model was developed in a training cohort that consisted of 501 patients. Radiomics features were extracted from the 18F-FDG PET/CT of the primary tumor. Support vector machine and extremely randomized trees were used to build the RM. Internal validation was assessed. An independent testing cohort contained the remaining 215 patients. The performances of the RM and clinical node staging (cN staging) in predicting pN staging (pN0 vs. pN1 and N2) were compared for each cohort. The area under the curve (AUC) of the receiver operating characteristic curve was applied to assess the model's performance. Results: The AUC of the RM [0.81 (95% CI, 0.771–0.848); sensitivity: 0.794; specificity: 0.704] for the predictive performance of pN1 and N2 was significantly better than that of cN in the training cohort [0.685 (95% CI, 0.644–0.728); sensitivity: 0.804; specificity: 0.568], (P-value = 8.29e-07, as assessed by the Delong test). In the testing cohort, the AUC of the RM [0.766 (95% CI, 0.702–0.830); sensitivity: 0.688; specificity: 0.704] was also significantly higher than that of cN [0.685 (95% CI, 0.619–0.747); sensitivity: 0.799; specificity: 0.568], (P = 0.0371, Delong test). Conclusions: The RM based on 18F-FDG PET/CT has a potential for the pN staging in patients with NSCLC, suggesting that therapeutic planning could be tailored according to the predictions.


INTRODUCTION
Among all cancers, lung cancer remains the most commonly diagnosed (11.6% of the total cases) and leading cause of cancer death (18.4% of the total cancer deaths). Non-small cell lung cancer (NSCLC) accounts for 85% of the cases (1,2). For patients newly diagnosed with NSCLC, the exact evaluation of the pathological lymph node (LN) status plays an important role in the choice of therapy regimen. There is a consensus that lobectomy combined with systemic nodal dissection is the recommended surgical treatment for early-stage NSCLC; however, sublobar resection and stereotactic body radiation therapy (SBRT) are possible alternatives for patients who are ineligible for lobectomy (3)(4)(5). Thus, accurate differentiation of pathological node-negative from positive is critical for selecting the optimal therapeutic plan.
Currently, one of the most widespread modalities used for the clinical LN (cN) staging of patients with NSCLC is fluorine-18fluorodeoxyglucose positron emission tomography/computed tomography ( 18 F-FDG PET/CT) (6,7). Unfortunately, the accuracy of 18 F-FDG PET for the direct evaluation of each mediastinal LN for the presence of metastasis is inherently limited by an avid FDG uptake that can be caused by inflammation due to infectious or non-infectious etiologies such as tuberculosis, pneumoconiosis, or chronic obstructive pulmonary disease (8)(9)(10). To improve the diagnostic ability of false-positive signs, several studies have analyzed the differences in parameters such as morphology, density, metabolism, and radiomics between benign and malignant LNs (11,12). Nevertheless, because of the low FDG uptake, occult LN metastasis (OLM) in patients with NSCLC fails to be detected by 18 F-FDG PET (13) and, hence, imaging is prone to false-negative signs. Accordingly, some researchers have tried to predict OLM by analyzing the 18 F-FDG metabolic parameters of the primary tumor in NSCLC (14,15). To the best of our knowledge, few researchers have dealt with both problems of false-positive signs and OLM in mediastinal LN staging. Furthermore, few studies have investigated whether radiomics features derived from the primary lesion of NSCLC might provide useful information for mediastinal LN staging. Therefore, we constructed and validated a radiomics model (RM) to predict the pathological mediastinal LN staging (pN0 vs. pN1 and pN2) based on the 18 F-FDG PET/CT imaging of NSCLC primary tumors.

Patients
This study involving human participants was reviewed and approved by the Ethical Commission of Medical Research Involving Human Subjects at the Region of Xiangya Hospital, Central South University, China, and the requirement for informed consent was waived. We reviewed the electronic medical records of 716 consecutive patients with NSCLC [adenocarcinoma (ADC) and squamous cell carcinoma (SCC)] who underwent both 18 F-FDG PET/CT staging and surgical resection with a curative intent from February 2007 to November 2019. All the patients underwent surgical resection with systematic mediastinal (N2) and hilar (N1) LN dissections within 2 weeks of 18 F-FDG PET/CT examination. Pre-operative cN staging and post-operative pN staging of the patients were performed and recorded according to the eighth edition of the Union for International Cancer Control TNM classification (16). Histological types were diagnosed according to the World Health Organization classification. We excluded patients from the study if they had (i) histology other than ADC or SCC, (ii) history of other cancer, (iii) received any treatment before 18 F-FDG PET/CT, and (iv) undergone pre-operative lung biopsy.

F-FDG PET/CT Acquisition and Reconstruction
All 18 F-FDG PET/CT scans were performed on a dedicated PET/CT scanner (Discovery ST8, GE Healthcare, Chicago, IL). All patients fasted for at least 6 h before imaging, and a blood glucose level of <110 mg/dL was confirmed before the administration of 18 F-FDG. PET/CT was performed ∼60 min after the intravenous injection of 370 MBq/kg of 18 F-FDG. First, a low-dose CT scan without contrast enhancement (120 mA, 150 kV, 512 × 512 matrix, the pitch of 1.75, reconstruction thickness and interval of 3.75 mm) for a precise anatomical localization and attenuation correction was performed. Next, a three-dimensional PET scan (thickness of 3.27 mm) was performed from the skull base to the proximal thighs with an acquisition time of 3 min per bed position.
The PET data sets were iteratively reconstructed using an ordered-subset expectation maximization (OSEM) algorithm with attenuation correction. All collected images were displayed on the GE Healthcare Xeleris 3.0 to reconstruct the PET, CT, and PET/CT fusion images.

Image Interpretation and Lesion Segment
Two experienced nuclear medicine physicians who were blinded to the patient's clinical information retrospectively reviewed the 18 F-FDG PET/CT scans. Any difference of opinion was resolved by consensus. Mediastinal and hilar LNs with a short axis of ≥10 mm in the short axis on CT and with a high accumulation of 18 F-FDG compared with that of the adjacent mediastinal tissue were considered as cN2 or cN1 at our institution. Fused PET/CT images were viewed on the Advantage Workstation (version AW 4.7, GE Healthcare).
The region of interest (ROI) for each patient was delineated initially around the tumor outline for the largest cross-sectional area of the primary lung lesion on both the CT and PET images. The ROIs were segmented manually by a single experienced nuclear medicine physician, and the final ROIs were checked by another nuclear medicine physician with more than 10 years of experience in PET/CT diagnosis. The open-source imaging platform ITK-SNAP software (version 3.6; www.itksnap.org) was used to plot the ROIs of the corresponding lesions (17). The feature data were extracted, pre-processed, modeled, evaluated, and validated using the scikit-learn (sklearn, scikit-learn.org) packages in the python platform (18).

Radiomics Feature Extraction
Data pre-processing: to ensure that the features were comparable, training/testing cohort division, missing-value filling, and data standardization were performed. First, to maintain the distribution of the original data, a stratified sampling method was applied to identify the training (501 samples, 70%) and testing cohorts (215 samples, 30%) (19). Moreover, the missing values (0 and 5) were filled with the median in the training and testing cohorts, respectively, and then, the same normalization was used for the data.
There were 1,438 features of primary tumors that were automatically extracted using the sklearn packages. The Spearman rank order correlation coefficient was used to calculate the relationship between features, and the redundant features were eliminated with an average absolute correlation of 0.85 as the threshold. Support vector machine-recursive feature elimination, and the extremely randomized trees were applied to reduce the dimensions and select optimized features for the radiomics model (RM) to avoid the impact of redundant and unconnected features. The relevance of the association between each radiomics feature was established using heat maps ( Figure 1). Consequently, a total of 25 principal correlative features, obtained through dimension reduction, were identified for inclusion in the RM to distinguish pN0 from pN1 and pN2. The results of the feature selection are shown in Table 1.

Radiomics Modeling and Evaluation
Extremely randomized trees was used as a classifier to model and optimize the radiomics signature in the modeling process. The 25 selected features were put into the classifier to build the RM to predict the pathological status of mediastinal LNs in the training cohort. Thereafter, a five-fold cross-validation of the training cohort was used to identify differences in the results. The training model was applied to the testing cohort for model validation.
The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was used as a means of quantitatively identifying the effective performance of the RM. The confusion matrix in the testing cohort was calculated (Figure 2).

Statistical Analysis
Statistical analyses were conducted using the SPSS software, version 23.0 (IBM Corm., Armonk, NY), and P-values < 0.05 were deemed statistically significant.
The predictive abilities of the RM and cN were investigated using ROC analysis. The statistical significance of the improvement in the AUC after adding an explanatory factor was evaluated using the Delong test (20).
The clinicopathologic characteristics of the patients with a pN0 status were compared with those of the patients with pN1 and pN2. The training and testing data cohorts were compared to identify factors contributing to nodal metastasis  Histogram feature Histogram parameters are related to the properties of individual pixels. They describe the distribution of voxel intensities within the images through the commonly used and basic metrics. Let X denote the 3D image matrix with voxels and the first-order histogram divided by discrete intensity levels.

Textural phenotype features
Texture is one of the important characteristics used in identifying objects or regions of interest in an image. Texture represents the appearance of the surface and how its elements are distributed. It is considered an important concept in machine vision; in a sense, it assists in predicting the feeling of the surface (e.g., smoothness, coarseness, etc.) from image.

PET_textural_phenotype_level_H
Intra-peri-nodular textural transition features Intra-peri-nodular textural transition features represents a minimal set of quantitative measurements which attempt to capture the transitional heterogeneity from the intrato the peri-nodular space.

PET_Ipris_shell0_ge_mean
Partial local pattern binary feature Partial local pattern binary feature is a local descriptor of the image based on the neighborhood for any given pixel. The neighborhood of a pixel is given in the form of P number of neighbors within a radius of R.

CT_PLBP_hist_tumor_orient1_2
High order texture feature based on wavelet transform By using a family of functions localized in terms of time and frequency, wavelet transforms can centralize the energy of the original image within only a few coefficients after wavelet decomposition. These coefficients have high local relativity in three directions of different sub-band images: horizontal, vertical, and diagonal.

Characteristics of All Patients
The clinicopathological characteristics of the 716 patients enrolled in the study are shown in Table 2. Figure 3 shows the patient recruitment pathway. Among them, 220 were female and 496 were male, with an age range from 25 to 78 years. ADC was the most common histological type of NSCLC (417/716). The number of patients with SCC was 329. In the training cohort, the number of patients with pN0, pN1, and pN2 were 315, 74, and 112, respectively. In the testing cohort, the number of patients with pN0, pN1, and pN2 were 135, 43, and 37, respectively. The age of the training cohort, sex of the testing cohort, and  anatomical classification of the two cohorts were statistically significantly different between the pN0 status and pN1 and pN2 status (P < 0.050). However, no significant difference was observed in the age of the testing cohort and in the sex of the training cohort (P > 0.050). Moreover, smoking history, lobar distribution, and histologic cell type were not significantly different between the two cohorts.

RM Performance
The diagnostic efficiency of the RM and cN were evaluated by the ROC curve.  The above-mentioned nuclear medicine physicians excluded the patients with LNs significant enlargement and intense 18 F-FDG uptake in PET/CT, and confirmed N1 or N2 by pathology from the whole population. The remaining 634 patients were defined as the cN ± group. Then, the sensitivity, specificity, and AUC of the cN and the RM in the cN ± group were calculated, respectively.
The RM showed the AUC of 0.802 (95%CI, 0.683-0.921) for the prediction of mediastinal LNs malignancy using the optimum cutoff value of 0.382 in the cN ± group. The sensitivity and specificity of the RM were 0.718 and 0.767, respectively. In comparison, the cN showed the AUC of 0.611 (95%CI, 0.572-0.650) using the optimum cutoff value of 0.500. The sensitivity and specificity of the cN were 0.819 and 0.404, respectively. The results demonstrated that the performance of the RM was effective in discriminating pN0 from pN1 and pN2 using the 18 F-FDG PET/CT images. The performances of the RM and cN in the training and testing cohorts and the cN ± group are displayed in detail in Figures 4,  5, respectively, and the representative cases are presented in Figure 6.

DISCUSSION
In cases of NSCLC with a chance of cure, the standard surgical procedure is a pulmonary lobectomy with systemic mediastinal nodal dissection. However, some patients are not eligible for this therapy because of their advanced age or the presence of severe medical diseases, and some patients refuse surgical treatment. Limited surgery (wedge resection or segmentectomy) or SBRT would be alternatives for such patients. Sublobar resection helps preserve more healthy lung tissue, shortens the operative time, and improves the post-operative quality of life. Perioperative mortality and operative complication morbidity do not differ significantly between lobar and sublobar resection (3). SBRT has emerged as the preferred management strategy for patients who are not surgical candidates; however, for the selection of SBRT or restrictive surgery, accurate prediction of a pathological LN-negative status is a pre-requisite.
The diagnosis of NSCLC mediastinal LN metastasis is generally based on several parameters such as metabolism, size, morphology, and attenuation, which leads to dependence on clinical experience. In other words, the traditional practice involves treating medical images as pictures intended solely for visual interpretation (21). False-positive findings of mediastinal LNs are not uncommon in functional imaging with 18 F-FDG PET/CT because the modality can mistakenly identify inflammation in patients with NSCLC due to infection, inflammation, or granulomatous diseases (8)(9)(10). The main molecular and pathological mechanisms of an avid FDG uptake in benign mediastinal LNs are lymphoid follicular hyperplasia and histiocyte infiltration associated with glucose transporter-1 overexpression (22). When benign mediastinal LNs manifest as a false-positive finding on PET imaging, and the CT morphology is not informative enough to support a judgment, there is an increased risk of an incorrect diagnosis. Benign high-uptake LNs can coexist with occult metastasis, making an accurate cN staging more difficult. In the clinical practice of mediastinal LN staging in NSCLC, nuclear medicine physicians are faced with the challenge of suspected positive LNs and possible OLM almost every day, which is difficult to deal with by relying solely on experience.
In the existing studies, for the accuracy of the cN staging in NSCLC, radiologists and nuclear physicians often analyzed parameters such as morphology and glucose metabolism or radiomic features of visible mediastinal LNs to improve the diagnostic ability of 18    analysis of parameters based on 18 F-FDG uptake have been proposed in light of these limitations. Ouyang et al. used the primary tumor-to-blood SUV ratio and metabolic parameters in clinical N0 lung ADC to predict OLM (14). Kim et al. investigated the OLM's predictability using SUV, metabolic tumor volume (MTV), and total lesion glycolysis (TLG) in patients with cN0 lung SCC before surgery (15). There are no current studies of 18 F-FDG PET/CT primary tumor-based FIGURE 6 | Patient 1: male, 55 years old, ADC, PET/CT showed lymph nodes in 4R (yellow arrow) with an intense FDG uptake, which was evaluated as cN2, predicted as N0 by the RM, and confirmed as pN0 after radical resection. The red arrow indicates the tumor in the right lower lobe (CT, PET) and the segmentation on ITK-SNAP. Patient 2: male, 66 years old, ADC, evaluated as cN0, but predicted as N+ by the RM, and confirmed as pN1 after radical resection. The black arrow indicates the primary lesion in the left upper lobe (CT, PET) and the segmentation on ITK-SNAP. radiomics classifiers of the LN staging (N0 vs. N1 and N2) in NSCLC. Therefore, if our solution proves to be feasible, it can be used to either differentiate benign and malignant LNs or determine OLM, thus leading to an informed therapeutic decision-making in the face of the challenge of false-positive and false-negative images.
Huang et al. have developed and validated a radiomics nomogram based on the primary tumor in a contrast-enhanced CT for pre-operative LN metastasis prediction in colorectal cancer (27). Inspired by their research achievement, we aim to introduce the radiomics modeling approach based on the 18 F-FDG PET-CT images of the primary lesion into the LN staging in NSCLC. On the basis of the radiomics hypothesis, intratumoral heterogeneity detected by imaging could be the expression of genomic heterogeneity, which implies a worse prognosis because tumors with more genomic heterogeneity are more likely to be resistant to treatment metastasis (28). Mediastinal LN staging in NSCLC is highly correlated with prognosis; therefore, we assumed that LN metastasis information may be obtained from intratumoral heterogeneity. Some studies had discovered pretherapy 18 F-FDG PET/CT or CT-based radiomics classifiers of survival or response in patients with NSCLC (29)(30)(31)(32). Therefore, it can obtain information from the primary lesion that is helpful for diagnosis or prognosis.
Nevertheless, the present study has several limitations too. Due to its retrospective design and performance at a single center, there is a risk of selection bias. A larger, multiinstitutional prospective randomized study is needed to confirm these results.

CONCLUSIONS
A Radiomics Model based on the 18 F-FDG PET/CT analysis provided useful information for mediastinal LN staging in patients with NSCLC. Therefore, therapeutic planning could be tailored according to predictions, and limited surgery or SBRT could be helpful in patients with cN0.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Ethical Commission of Medical Research Involving Human Subjects at Region of Xiangya Hospital, Central South University, China. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements. Written informed consent was not obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

AUTHOR CONTRIBUTIONS
KZ designed the method, acquired the data, and wrote and edited the article. SH designed the method and approved the manuscript. Image analysis was performed by XW. Statistical analysis was performed by XW and CJ. YT, ZF, JH, and ZZ aided in data acquisition and interpretation. All authors contributed to the article and approved the submitted version.

FUNDING
This study was supported by the National Natural Science Foundation of China (Nos. 91859207, 81771873, and 81471689).