Intratumoral and Peritumoral CT-Based Radiomics Strategy Reveals Distinct Subtypes of Nonsmall-Cell Lung Cancer

Purpose To evaluate a new radiomics strategy that incorporates peritumoral and intratumoral features extracted from lung CT images with ensemble learning for pretreatment prediction of lung squamous cell carcinoma (LUSC) and lung adenocarcinoma (LUAD). Methods A total of 105 patients (47 LUSC and 58 LUAD) with pretherapy CT scans were involved in this retrospective study and were divided into training (n=73) and testing (n=32) cohorts. Seven categories of radiomics features involving 3078 metrics in total, were extracted from the intra- and peritumoral regions of each patient’s CT data. Student’s t-tests in combination with three feature selection methods were adopted for optimal features selection. An ensemble classier that was generated with ve machine learning classiers and optimal features, was developed and the performance was quantitatively evaluated using both training and testing cohorts for the prediction task. Results The classication models developed by using optimal feature subsets determined from intratumoral region and peritumoral region with the ensemble classier achieved mean area under the curve (AUC) of 0.87, 0.83 in the training cohort and 0.66, 0.60 in the testing cohort, respectively. The model developed by using the optimal feature subset selected from both intra- and peritumoral regions with the ensemble classier achieved great performance improvement, with AUC of 0.87 and 0.78 in both cohorts, respectively. Conclusions The proposed new radiomics strategy that extracts image features from the intra- and peritumoral regions with ensemble learning, could greatly improve the diagnostic performance for the histological subtype stratication in patients with NSCLC.


Introduction
Lung cancer is the most frequently occurring cancer and the leading cause of cancer-related death in men globally [1]. In women, lung cancer is the third most commonly diagnosed cancer and the second most leading cause of cancer-related death [1]. Approximately 85% of primary lung malignancies are nonsmallcell lung cancer (NSCLC), and the 5-year survival rate is less than 20% [2][3][4][5][6].
The rst-line reference in preoperatively diagnosing LUSC and LUAD is lung biopsy [3,[6][7][8]10], which is an invasive diagnostic approach with a high level of risks in clinical practice [11]. In addition, concerning the issue of tumor heterogeneity of NSCLC, lung biopsy examines only very limited proportions of the tumor tissue and is incapable of completely characterizing tumor properties [5,7]. Developing a noninvasive strategy for the accurate prediction of LUSC and LUAD preoperatively is desirable.
Noninvasive imaging technologies, such as computed tomography (CT) and multiparametric magnetic resonance imaging (mpMRI), have recently been widely used for the pretherapy diagnosis of NSCLC [5,7,[12][13][14][15]. Compared with mpMRI, CT offers considerably better imaging e ciency, higher resolution, and fewer motion artifacts caused by breathing and is thus recommended in the guidelines for NSCLC screening and diagnosis [2,13]. However, it is very challenging for clinicians to visually predict the histological subtype of NSCLC directly from CT images to discriminate between LUSC and LUAD.
In recent years, radiomics strategies have been used for the prediction of LUSC and LUAD. In 2016, Wu  Chaunzwa et al. introduced the convolutional neural network (CNN) to the prediction task and developed a prediction model based on the Visual Geometry Group-16 (VGG-16) network [17], obtaining an optimal AUC of 0.751.
In addition, some recent studies also integrated the radiomics strategy with positron emission tomography computed tomography (PET-CT) images, achieving favorable diagnostic performance in the differentiation of these two subtypes of NSCLC [18][19][20]. For instance, Koyasu et al. proposed a PET-CTbased radiomics strategy with an extreme gradient boosting (XGBoost) classi er for the prediction task [19], achieving good performance with an AUC of 0.843.
Although these previous studies have repeatedly demonstrated the feasibility of the radiomics strategy based on CT or PET-CT for the prediction of histological subtypes of NSCLC, all the features they extracted were from the intratumoral region of the image. We are not aware of any work that has attempted to evaluate the peritumoral area outside the tumor to distinguish LUSC from LUAD. According to a recent study [21], perinodular region-based radiomics features on lung CT images effectively re ect the difference between LUAD and granulomas and accurately distinguish these two types of lung nodules. Whether the radiomics features extracted from the peritumoral region of NSCLC can re ect the signi cant difference between LUSC and LUAD and further be used for the prediction task remains an open question to date. Therefore, the rst aim of this study was to investigate whether the radiomics features extracted from the peritumoral region of NSCLC could signi cantly re ect the difference between LUSC and LUAD. To achieve this goal, seven feature categories were employed in this study, including morphological features, histogram-based features ( rst-order features, hereafter), Haralick features of co-occurrence matrices (CM features, hereafter) [22], and features derived from the run length matrix (RLM features, hereafter) [23], the neighborhood gray-tone difference matrix (NGTDM features, hereafter) [24], the gray level size zone matrix (GLSZM features, hereafter) [25], and gray level dependence matrix (GLDM features, hereafter) [26] to fully characterize the global, local and regional differences of the tissue in the peritumoral region between LUSC and LUAD [27].
The second aim was to develop an accurate and consistent model for predicting LUSC and LUAD. To ful l this aim, both intra-and peritumoral region-based radiomics features were utilized, and an ensemble classi er that combined multiple binary classi ers, such as support vector machine (SVM), RF and XGBoost, was used to form a more robust predictive model. The diagnostic performance of the model was then assessed with AUC for the differentiation of LUSC and LUAD.

Materials And Methods
This retrospective study was approved by the institutional ethics review board of Xijing Hospital, and informed content was waived. The overall methodological pipeline of this study is shown in Fig. 1.

Patients
A total of 146 archival patients with postoperatively con rmed NSCLC were collected from Xijing Hospital. The inclusion criteria were as follows: i) primary LUSC or LUAD was pathologically con rmed; ii) CT scan was performed prior to any therapies. Patients who met one of the following conditions were excluded: i) lack of postoperative pathological information to con rm the histopathological subtype of the patient as LUSC or LUAD (n=21); ii) missing preoperative CT scan (n=16); or iii) poor imaging quality makes accurate tumor annotations extremely di cult (n=4). Finally, 105 subjects were eligible for this study, including 47 patients with LUSC and 58 patients with LUAD. The patients were then randomly allocated into the training cohort (n=73) and testing cohort (n=32). The inclusion-exclusion process is illustrated in Fig. 2.

Image acquisition and region of interest annotation
All patients underwent thoracic CT imaging using a uCT 760 system (United Imaging Healthcare, China). The primary scanning parameters were as follows: 80 kV; 80 mAs; detector collimation: 64 × 0.6 mm; rotation time: 0.4 s; slice thickness: 5 mm; spacing between slices:5 mm; pixel spacing: 0.6 × 0.6 mm; and matrix size, 512 × 512. The entire lung region was scanned in each patient, and the image slice varied from 100 to 400.
Two types of regions of interest (ROIs) including intra-and peritumoral regions, were annotated from the CT images, as shown in Fig. 3. Prior to the intratumoral region annotation of each CT dataset, the axial image slice was selected to obtain the largest area of the archived tumor with the maximal size in each patient's lung region. Then, a manually depicted polygonal ROI was used to segment the intratumor region on the selected image slice. Two radiologists with 20 and 10 years of lung CT interpretation experience independently performed intratumoral region delineation using a custom-developed package. Then, divergence of their delineation results was carefully corrected by consensus.
After the intratumoral region mask was obtained, we adopted the morphological dilation operator to generate a new region mask that was approximately 10 mm larger in radial distance than the intratumoral region according to pixel size [21]. Then, the corresponding peritumoral region was the ring of the lung parenchyma around the tumor that was obtained by subtracting the intratumoral region mask from the new region mask after morphological expansion, as shown in Fig. 3. Finally, the peritumoral region was further divided into two rings including the rst ring (0-5 mm) and the second ring (5-10 mm) for feature extraction and comparison [21].

Radiomics feature extraction
After intra-and peritumoral ROI segmentation, 10 lters including wavelet-HL, wavelet-LL, wavelet-LH, wavelet-HH, square, square root, logarithm, exponential, gradient, and local binary pattern (LBP), were utilized to the original image to magnify the tissue patterns and unearth important features. Then, six feature categories, including rst-order features, GLCM features, GLRLM features, NGTDM features, GLSZM features and GLDM, were calculated from the original segmented image data and 10 generated images of the intratumoral and two rings of the peritumoral regions [28]. Given that the peritumoral region was dilated based on use of the intratumoral region, the shape 2D features were only calculated from the intratumoral region. Therefore, 1032, 1023, 1023 radiomics features were extracted from the intratumoral region and the rst ring and the second ring of the peritumoral region, respectively, as shown in Table 1. Open source Pyradiomics (version 3.0.1) was used to perform this analysis [29]. All of the codes and results have been attached in the Appendix document. In this study, a two-step feature selection strategy was adopted to determine an optimal subset of features for model construction, as shown in Fig. 1. The rst step was statistical analysis of all these features between LUSC and LUAD, which was performed with Scikit-learn. Student's t-test with a signi cant p-value set as 0.05 was then performed with all radiomics features to select those with signi cant intergroup differences between LUSC and LUAD [30].
Then, all signi cant features were standardized to eradicate differences of the feature-value scales. The normalized feature of each feature for a speci c patient is calculated as follows: (1) where and are the mean and standard deviation, respectively, of each feature from the training cohort.
In the second step of feature selection, three widely-used feature selection algorithms, including the minimum redundancy maximum relevance method (mRMR) [31], the least absolute shrinkage and selection operator(LASSO) [32,33], and the linear SVM-based recursive feature elimination (SVM-RFE) [34], were further implemented with these signi cant features to select an optimal feature subset from the training cohort for model development and external testing.

Model development based on ensemble learning and validation
With optimal features selected, the predictive model was developed using the training cohort and the ensemble learning strategy, which includes ve commonly used binary classi ers, including the quadratic discriminant analysis(QDA) classi er, SVM with radial basis function(RBF) kernel, SVM with sigmoid/tanh kernel, RF, and XGBoost. QDA is the most commonly used binary classi er, which has no same-covariance assumption for each binary class [35,36]. SVM is a classical machine learning classi er with several typical kernels, such as RBF and sigmoid/tanh, that is used to compute the decision boundary that separates two classes with the maximum marginal distance [37][38][39]. It has advantages in dealing with nonlinear features and is not easily over t with even small datasets [40]. The RF classi er can build multiple random decision trees (100 trees of the default parameter in Scikit-learn to avoid over tting) and integrate them to make an accurate diagnosis [40][41][42]. XGBoost offers many bene ts in classi cation, including high precision and consistency and the prevention of over tting [43,44]; thus, it was also included in the ensemble learning strategy.
The ensemble classi er was nally developed by weighting the predictive value of these ve classi ers in the model training process, which can be expressed as follows: (2) where represents the nal predictive value of the j-th patient; denotes the predictive value of the j-th patient by using the i-th classi er; and is the weighting parameter of the i-th classi er in the ensemble learning process, which meets the following condition: (3) In this study, the optimal weight was determined based on minimizing the predictive error in the training process, and the cutoff for assigning the patient to the LUAD group was set as 0.5. If was greater than or equal to 0.5, the j-th patient was allocated to the LUAD group. The overall performance was evaluated using both the training cohort and the testing cohort with the quantitative metric of AUC [45][46][47][48]. The AUC value was widely used to comprehensively evaluate the general performance of the model developed for the prediction task [45][46][47][48]].

Statistical analysis
Statistical analyses of the patient demographics were performed using IBM SPSS statistics (version 19.0, Armonk, NY), and Python software (version 3.6 DL-GPU) was used to perform statistical selection of features with signi cant differences between LUSC and LUAD. Chi square tests were performed to evaluate signi cant differences in primary clinical factors distributed between the training and testing cohorts, and Student's t-tests were used to select signi cant radiomics features between LUSC and LUAD. Two-sided p-values less than 0.05 were considered signi cant [27,49,50].

Demographics of eligible patients
A total of 105 NSCLC patients were eligible for this study, including 47 patients with LUSC and 58 with LUAD. These patients were randomly allocated into the training cohort (n=73) and the testing cohort (n=32). The baseline demographics and clinical information of these patients was collected from the archival medical document, as shown in Table 1. Statistical analyses indicate no signi cant differences between both the training and testing cohorts in terms of all these primary factors.
3.2 Results of the two-step feature selection strategy A total of 3078 standardized radiomics features, including 1032 features from the intratumoral region, 1023 from the rst ring (0-5 mm) and 1023 from the second ring (5-10 mm) of peritumoral regions, were analyzed using Student's t-test (p-value < 0.05) to determine those with signi cant intergroup differences between LUSC and LUAD. Eventually, 500 signi cant features were selected from the intratumoral region, whereas, only 220 and 119 signi cant features were selected from the rst ring and second ring of peritumoral regions, respectively, as shown in Fig. 4. These results indicate that i) a large number of radiomics features extracted from the peritumoral region can also re ect the signi cant differences in tissue distribution patterns between LUSC and LUAD; ii) the closer the peritumoral region is located to the intratumoral region, the more features with signi cant differences could be obtained to re ect the tumor property difference. Fig. 5 illustrates an example of the intra-and peritumoral tissue distribution differences of LUSC and LUAD determined using one of the signi cant radiomics features, energy, with 3×3 sliding patches on the CT image. After statistical analysis-based feature selection, three radiomics feature subsets were nally obtained, including i) 500 signi cant features from the intratumoral region, ii) 339 signi cant features from the entire peritumoral region, and iii) 839 signi cant features from both intratumoral and peritumoral regions. All these signi cant features in each feature subset were further selected using three commonly applied strategies: SVM-RFE, LASSO, and mRMR with the mutual information difference (MID), as shown in Figs.
6 -8. Table 2 shows the results after the second-step feature selection procedure.

Classi cation model development and performance evaluation
As these optimal feature subsets were determined, classi cation models were developed using ve commonly used machine learning classi ers and the ensemble classi er with the training cohort, and the performance of each model was evaluated by using both training and testing cohorts for distinguishing LUSC from LUAD. The results are presented in Fig. 9. Three columns of sub gures in Fig. 9 exhibit the performance of predictive models developed by using optimal feature subsets determined from the intratumoral region, peritumoral region, and both intra-and peritumoral regions. These ndings indicate that i) the classi cation model determined from the peritumoral region achieved comparable performance to that from the intratumoral region; ii) the classi cation model determined from intra-and peritumoral regions dramatically improved the overall performance for the prediction of LUSC and LUAD; and 3) the model developed by the ensemble classi er achieved more favorable and consistent performance with training and testing cohorts compared with those developed by ve independent classi ers. Table 3 shows the performance of classi cation models developed by the ensemble classi er for the prediction task, indicating that the ensemble classi cation model developed by SVM-RFE-based optimal features determined from intra-and peritumoral regions achieved the best performance with AUC values of 0.87 and 0.78 in the training and testing cohorts, respectively.

Discussion
In this study, we investigated the feasibility of CT-based radiomic features extracted from intra-and peritumoral regions of NSCLC to re ect the tissue distribution differences between LUSC and LUAD, and developed a CT-based radiomics strategy that incorporated high-throughput features with an ensemble classi er for the preoperative prediction of LUSC and LUAD. Three widely used methods, SVM-RFE, LASSO, and mRMR, were employed to select optimal features with signi cant intergroup differences between LUSC and LUAD for classi cation model development. Five independent classi ers, QDA, SVM with RBF kernel, SVM with sigmoid/tanh kernel, RF, and XGBoost, which were reported to have favorable classi cation performance and robustness for the diagnosis of cancer phenotypes with a small database, were utilized to form an ensemble classi er for classi cation model building. The results of the model that was developed using the ensemble classi er and optimal features selected by SVM-RFE from intra-and peritumoral regions demonstrate favorable discriminative power with both the training and testing cohorts.
In recent years, CT-/PET-CT/multimodal MRI-based radiomics strategies have been repeatedly demonstrated to have great capability for the prediction of LUSC and LUAD [2,9,16,[18][19][20]. The diagnostic performance ranged between 0.72 and 0.843. Nevertheless, all these previous studies only focused on how to extract an increasing number of features from the intratumoral region of the image, regardless of the peritumoral parenchyma, which might also contain substantial information and be of equal importance for the prediction task. Some studies have revealed that the interface of the tumor has a "rim" of densely packed tumor-in ltrating lymphocytes and tumor-associated macrophages in representative hematoxylin and eosin-stained images [8, 21,51,52]. At a macroscopic scale, the densely packed stromal tumor-in ltrating lymphocytes around LUAD represent ne and smooth textures on CT images and thus could be potential imaging biomarkers for the identi cation of LUAD from LUSC [21]. However, whether radiomics features extracted from the peritumoral parenchyma region effectively re ect the intergroup difference of the tissue and microenvironment between LUSC and LUAD, remains unknown to date.
In this study, we found that a large number of radiomics features extracted from the intratumoral region and peritumoral region were signi cantly different between LUSC and LUAD, and the total number of signi cant features extracted from the rst ring (0-5 mm) peritumoral region was much greater than that of the signi cant features extracted from the second ring (5-10 mm) peritumoral region. These results demonstrate and verify for the rst time the hypothesis that the peritumoral region on CT images also contains substantial information that can re ect the tissue texture difference between LUSC and LUAD. In addition, the closer the peritumoral region is to the intratumoral region, the more substantial the information it contains.
Most of the previous studies only focused on extracting features from the original image data, neglecting the image lters that not only reduce the noise but also enhance the quality and magnify the texture in the image [53,54]. Therefore, in this study, 10 lters including wavelet-HL, wavelet-LL, wavelet-LH, wavelet-HH, square, square root, logarithm, exponential, gradient, and LBP were utilized to preprocess the image for feature extraction. Seven categories of radiomics features, including morphological features, rst-order features, second-order features, and high-order texture features, were adopted in this study to fully characterize the shape properties and global, local and regional distribution patterns of the tissue, respectively. Student's t-tests integrated with three widely applied feature selection algorithms (SVM-RFE, LASSO and mRMR), were adopted for optimal feature selection and performance comparison. The results indicate that the optimal features selected using the SVM-RFE algorithm from all signi cant features of both intra-and peritumoral regions have the most powerful diagnostic ability for the discrimination between LUSC and LUAD.
Classi cation model development is the last but most crucial step in the proposed radiomics strategy for the prediction of LUSC and LUAD. In this step, the choice of an optimal decision classi er, for instance, SVM with RBF kernel or Sigmiod kernel, RF, QDA, or XGBoost represent the core in uence of performance variation [40]. Hence, the determination of an optimal classi er is of critical importance. To fully integrate all the merits of these ve independent classi ers, an ensemble classi er was generated using ve independent classi ers, SVM with RBF kernel, or sigmoid kernel, RF, QDA, and XGBoost, and its diagnostic performance was compared with these independent classi ers. The results indicate that i) the classi cation model developed using the ensemble classi er achieves the most favorable, consistent and robust diagnostic performance compared with other independent classi ers, and ii) optimal features determined by SVM-RFE from both intra-and peritumoral regions with the ensemble classi er achieve the best diagnostic performance for the prediction of LUSC and LUAD with both training and testing cohorts. In addition, the classi cation results of all these models developed by each classi er with optimal features determined from intratumoral, peritumoral, or both of intratumoral and peritumoral regions using SVM-RFE, LASSO, and mRMR also revealed that although the model based on the ensemble classi er did not always obtain the best results, it always ranked as one of the top two models in terms of the AUC with both cohorts, suggesting remarkable consistency and robustness in the prediction of LUSC and LUAD.
The limitations of this study include the following aspects. First, inherent bias might exist given the retrospective nature of the present study with relatively small patient cohorts collected from a single clinical center. A larger number of participants from two or more clinical centers are further required to validate the performance of the model we developed. Moreover, other potential clinical factors, such as gene mutations and key molecular biomarkers, were not included in the current study given the incomplete data in the archival database, which should be further analyzed. In addition, deep radiomics features incorporating the current manual radiomics features might further improve current performance in the prediction of LUSC and LUAD.
In conclusion, the proposed CT-based radiomics strategy that extracts features from intra-and peritumoral regions, adopts SVM-RFE for optimal feature selection, and utilizes ensemble learning for classi cation model development is demonstrated with favorable predictive precision and stability for preoperatively prediction of LUSC and LUAD.

Con ict of Interest
The authors declare that they have no con ict of interest.

Ethics Approval
This study was approved by the institutional ethics review board of Xijing Hospital, and informed content was waived.

Data Availability Statement
The raw/processed data of this study cannot be publicly shared at present as it forms part of an ongoing study, but it could be available under reasonable request from the corresponding author with the permission of the Institutional Review Board. Results and code package in each step of this study have been arranged in a document named as "Appendix". The code package has also been uploaded to Gitee for publicly sharing and further perfection (https://gitee.com/yang-tianran-01/radiomics_-  The schematic pipeline of the proposed strategy for the prediction of lung squamous cell carcinoma (LUSC) and lung adenocarcinoma (LUAD) via intra-and peritumoral CT radiomics features and ensemble learning Figure 2 Inclusion-exclusion criteria of this study to obtain 105 eligible subjects including 47 ones with lung squamous cell carcinoma (LUSC) and 58 with lung adenocarcinoma (LUAD) Figure 3 Illustration of the intratumoral region (light green) manually delineated and the rst ring (0 -5 mm, light purple) and second ring (5 -10 mm, red) of the peritumoral regions generated by morphologically expanding the segmented intratumoral region mask  Intra-and peritumoral tissue distribution differences between LUSC and LUAD characterized by the signi cant radiomics feature Energy on CT images with the unit normalized as "1" on the color bar Figure 6 Optimal features selected using SVM-RFE approach: (a) 12 optimal features selected from the intratumoral region; (b) six optimal features selected from the peritumoral region; and (c) nine optimal features selected from intra-and peritumoral regions.

Figure 7
Optimal features selected using LASSO approach: (a) six optimal features selected from the intratumoral region; (b) six optimal features selected from the peritumoral region; and (c) eight optimal features selected from intra-and peritumoral regions. Optimal features selected using mRMR with MID: (a) 12 optimal features selected from the intratumoral region; (b) 12 optimal features selected from the peritumoral region; and (c) 12 optimal features selected from intra-and peritumoral regions. Figure 9