Evaluating Histological Subtypes Classification of Primary Lung Cancers on Unenhanced Computed Tomography Based on Random Forest Model

Lung cancer is the leading cause of cancer-related death in many countries, and an accurate histopathological diagnosis is of great importance in subsequent treatment. The aim of this study was to establish the random forest (RF) model based on radiomic features to automatically classify and predict lung adenocarcinoma (ADC), lung squamous cell carcinoma (SCC), and small cell lung cancer (SCLC) on unenhanced computed tomography (CT) images. Eight hundred and fifty-two patients (mean age: 61.4, range: 29–87, male/female: 536/316) with preoperative unenhanced CT and postoperative histopathologically confirmed primary lung cancers, including 525 patients with ADC, 161 patients with SCC, and 166 patients with SCLC, were included in this retrospective study. Radiomic features were extracted, selected, and then used to establish the RF classification model to analyse and classify primary lung cancers into three subtypes, including ADC, SCC, and SCLC according to histopathological results. The training (446 ADC, 137 SCC, and 141 SCLC) and testing cohorts (79 ADC, 24 SCC, and 25 SCLC) accounted for 85% and 15% of the whole datasets, respectively. The prediction performance of the RF classification model was evaluated by F1 scores and the receiver operating characteristic (ROC) curve. On the testing cohort, the areas under the ROC curve (AUC) of the RF model in classifying ADC, SCC, and SCLC were 0.74, 0.77, and 0.88, respectively. The F1 scores achieved 0.80, 0.40, and 0.73 in ADC, SCC, and SCLC, respectively, and the weighted average F1 score was 0.71. In addition, for the RF classification model, the precisions were 0.72, 0.64, and 0.70; the recalls were 0.86, 0.29, and 0.76; and the specificities were 0.55, 0.96, and 0.92 in ADC, SCC, and SCLC. The primary lung cancers were feasibly and effectively classified into ADC, SCC, and SCLC based on the combination of RF classification model and radiomic features, which has the potential for noninvasive predicting histological subtypes of primary lung cancers.


Introduction
Random forest (RF) algorithm, proposed by Leo Breiman [1], is an ensemble learning algorithm based on classifcation and regression trees (CART). Te RF algorithm contains several CARTs, and each one is independent. Terefore, the RF algorithm performs insensitively to the overftting problem of the training cohort and has superior noise immunity, which is not sensitive to default values [2]. Te RF algorithm is widely applied in various felds, such as life sciences because RF classifcation models are versatile, have high prediction accuracy, and provide additional information such as variable importance [3]. According to literature reports, RF algorithms have excellent performance in evaluating the progression, prognosis, and gene mutation expression of various diseases [4][5][6][7][8][9][10].
Lung cancer is the leading cause of cancer-related death in many countries, and the accurate histopathological diagnosis is of great importance in subsequent treatment [11,12]. In previous studies, the RF model was mostly applied to detecting lung cancer, the classifcation of benign and malignant pulmonary nodules, and the analysis of lung cancer prognosis [13][14][15][16]. However, for therapeutic purposes, primary lung cancers fall into three major subtypes: lung adenocarcinoma (ADC), lung squamous cell carcinoma (SCC), and small cell lung cancer (SCLC), and distinguishing among subtypes is still particularly challenging. In this study, our aim is to establish a classifcation model combining the RF algorithm and radiomic features of unenhanced CT images to classify the primary lung cancers into ADC, SCC, and SCLC and to evaluate the prediction performance.

Study Population.
Nine hundred and twenty patients with histopathologically confrmed primary lung cancer from January 2013 to August 2018 at Zhongshan Hospital, Fudan University, were retrospectively studied. Te inclusion criteria were: (1) diagnosis of ADC, SCC, or SCLC confrmed by puncture or surgical specimen; (2) preoperative CT examination within 2 weeks before surgery. Te exclusion criteria were: (1) patients receiving other treatments such as chemotherapy and radiotherapy before surgery; (2) patients with lesion boundaries that were difcult to identify on CT images; (3) patients with inadequate quality images on CT; (4) patients with two or more histopathological subtypes of primary lung cancer; and (5) patients with the lesion less than 1 cm in diameter, avoiding partial volume efects. Eight hundred and fftytwo patients with primary lung cancer (525 ADC, 161 SCC, and 148 SCLC) were ultimately included in this study. All study procedures were approved by the Ethics Committee of Zhongshan Hospital, Fudan University.

Protocol of Unenhanced Computed Tomography and
Segmentation. All patients had preoperative CT examinations performed within 2 weeks before the puncture or surgery under breath-hold conditions at the end of inspiration, from the thoracic inlet to the diaphragm, by experienced radiologists. Te parameters of μCT 760 (Shanghai United Imaging Healthcare) were: tube voltage � 120 kV, tube current � 130 mA, slice thickness � 1 mm, and the parameters of LightSpeed 16 (GE Healthcare) were: tube voltage � 120∼140 kV, tube current � 140 mA, and slice thickness � 1 mm. All image data were stored in DICOM format.
ITK-Snap software (version 3.6.0) was used to segment each layer of the tumor lesions on the CT images in all cases to obtain a three-dimensional region of interest (ROI) [17], which was output in mha format for analysis (see Figure 1). Te histopathological results (ADC, SCC, or SCLC) of each case were matched to the segmentation results. All procedures of ROI segmentation were performed by two experienced radiologists and fnally confrmed by a senior radiologist.

Extraction of Radiomic
Features. PyRadiomics package implemented in Python was used to extract radiomic features [18] (see Figure 2). Te radiomic features of both the original and wavelet denoised (db2 was set as the wavelet basis) images were extracted, such as shape-based features, frstorder statistics, the gray level co-occurrence matrix (GLCM), the gray level run length matrix (GLRLM), the neighboring gray tone diference matrix (NGTDM), the gray level dependence matrix (GLDM), and the gray level size zone matrix (GLSZM) (see Figure 2). Te features were frst normalized to the range 0-1, and then the support vector machine (SVM) was used to flter the features (see Figure 2). Te variance infation factor (VIF) was used to detect the collinearity of features which made the area under the receiver operating characteristic (ROC) curve (AUC) of SVM classifcation greater than 0.5, and features with VIF less than or equal to 5 were selected (see Figures 2 and 3). Te formula of VIF is given as follows: (1) Te spatially uniform relevant features (ReliefF) algorithm was used to further flter the features (see Figure 2), and the fnal retained radiomic features were summarized in Table 1 (see Table 1). Te importance score of each feature for predicting the histopathological subtypes of primary lung cancer are shown in Figure 4 (see Figure 4).

Random Forest Algorithm.
Selected radiomic features were used to establish and RF classifcation model with the following parameters: "n_estimators" = 100; "max_depth" = 11; "min_samples_split" = 2; and "min_-samples_leaf" = 4. In this study, 85% of the whole data (724 in all, 446 ADC, 137 SCC, and 141 SCLC, respectively) were randomly divided into the training cohort which was used for feature selection as well as model ftting, and 5-fold crossvalidation was used to validate in the training cohort, while the remaining 15% (128 in all, 79 ADC, 24 SCC, and 25 SCLC, respectively) were divided into the testing cohort for validation (see Table 2 and Figure 5).

Statistical Analysis.
Statistical analysis was performed using SPSS software (Version 22.0) and Python 3.8.0 (NumPy packages). Categorical variables were presented as quantities (percentages) and compared using the chi-square test or Fisher's exact test, while continuous variables were presented as the mean ± SD if normally distributed, and compared using the Kruskal-Wallis H test because of the heterogeneity of variance. AUC, sensitivity, specifcity, and accuracy were used to evaluate the predictive performance of the classifcation model. In addition, the F1 score was also used to evaluate the efciency of the classifcation models. A two-tailed p value <0.05 was considered statistically signifcant. Table 3 (see Table 3 Table 2).  Notably, the diferences in age, gender, and TMN stage of patients among the three subtypes were statistically signifcant.

Predictive Performance of Random Forest Classifcation
Model. Twenty radiomic features were ultimately selected after features were extracted and fltered from the unenhanced CT to establish the RF classifcation model, including 7 frstorder features, 3 GLSZM features, 2 GLRLM features, 4 GLDM features, and 4 NGTDM features.

Discussion
Te histopathological diagnosis and classifcation of primary lung cancers are of great importance and crucial clinical value for the decision of optimal and individualized treatment schedules and the evaluation of prognosis [19]. In this study, RF algorithms combined with radiomic features on unenhanced CT images were used for noninvasive and preoperative prediction of subtype's classifcation of primary lung cancer. Radiomic features were extracted and fltered from enhanced CT images, and the ultimately 20 selected features were used to establish the RF classifcation model, which was trained and validated using the training cohort and the testing cohort. To be noted, 5-fold cross-validation was used for more accurate precision. Finally, the prediction performance of the model in classifying the three major subtypes (ADC, SCC, and SCLC) of primary lung cancers was evaluated. Te results showed that the RF classifcation model was able to accurately classify the three subtypes on the testing cohort (AUC � 0.80). Particularly, the model performed better in predicting SCLC (AUC � 0.88) than ADC (AUC � 0.74) and SCC (AUC � 0.77). However, this model tended to misclassify SCC as ADC, thus the recalls of the RF model in ADC (0.86) and SCLC (0.76) were excellent, while inferior in SCC (0.29). It was probably because (1) the sample of SCC (161) was limited and much fewer than that of ADC (525) and (2) most of the SCC included in this study were central-type lung cancer, which was difcult to distinguish on the CT images, leading to inaccurate segmentation of ROI. Certainly, the reason for the misclassifcation deserved further investigation and verifcation. In this study, the selected 20 radiomic features were not the same as the features in previous studies (com_radNet model), but they improved the predictive classifcation of SCLC [20].
Previously, a large number of studies have proven the excellent performance of the RF algorithm in classifying benign and malignant pulmonary nodules on CT and PET/ CT [21][22][23][24]. Zhu et al. classifed ADC and SCC in 129 patients with non-SCLC (NSCLC) based on 5 radiomic features with an AUC, specifcity, and sensitivity of 0.89, 0.90, and 0.83, respectively, on the validation cohort [25]. Liu et al. classifed 349 patients with NSCLC, including not only ADC and SCC but also large cell carcinoma and not otherwise specifed based on radiomic features combined with SVM, and the classifcation accuracy was 0.86 on the testing set [26]. In this study, we expanded the samples and    also investigated the classifcation between SCLC and NSCLC, with the considerably improved predictive performance of the RF classifcation model. To our knowledge, the only radiomics-based study on the classifcation of SCLC and NSCL identifed ADC, SCC, and SCLC in a two-by-two comparison. Te results showed good classifcation performance between ADC and SCLC (AUC � 0.86) and between ADC and SCC (AUC � 0.80) on unenhanced CT images and better performance on enhanced CT, but neither could efectively classify SCC and SCLC (AUC � 0.62 and 0.66). To note, the RF classifcation model was able to classify SCLC with great performance in our study. Te study has some limitations. First, the sample of ADC was much larger than that of SCC and SCLC, mainly due to the diferent morbidities, which may afect the diagnostic performance of the model [27]. Furthermore, although we have excluded cases with blurred tumor borders, the possibility of missegmentation of nontumor tissues existed. Finally, large multicenter, prospective studies are essential for model expansion and optimization.

Conclusions
In conclusion, the noninvasive histopathological subtype classifcation of primary lung cancers has great clinical signifcance and value. In our study, the primary lung cancers were feasibly and efectively classifed into ADC, SCC, and SCLC based on the combination of the RF classifcation model and radiomic features. Large studies are needed to optimize and validate the performance of the model. Te RF classifcation model combined with radiomic features on unenhanced CT images is able to provide additional information about patients and has the potential for clinical applications.

Data Availability
Te datasets used and/or analysed during the current study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that there are no conficts of interest.