Introduction

Non-small cell lung cancer (NSCLC) accounts for approximately 85% of lung cancer that is the most common cause of cancer-related mortality worldwide, with an estimated 1.4 million deaths each year [1]. Adenocarcinoma (ADC) and squamous cell carcinoma (SCC) are the most common subtypes of NSCLC [2]. Different pathological subtypes have distinct phenotypic and biological characteristics, which are directly related to the clinical treatment and outcome [3,4,5]. With advances in targeted therapies, molecularly targeted agents that inhibit epidermal growth factor receptor (EGFR) and anaplastic lymphoma kinase (ALK) can significantly improve the efficacy and reduce the toxicity of NSCLC, as almost all these gene mutations are found in ADC [6, 7]. Therefore, accurately predicting the histological subtypes is essential for determining better therapeutic strategies in NSCLC.

An invasive biopsy for histological confirmation is commonly used in clinical practice [8]. However, with the development of various detection technologies in recent years, high-precision noninvasive detection has been paid more attention and recognized by clinicians; moreover, biopsy is contraindicated for patients with severe cardiopulmonary insufficiency, such as severe pulmonary arterial hypertension, or uncorrectable coagulopathy, or unable to cooperate with the operation [9, 10]. In addition, when the pathological tissue obtained from the first puncture is few and fails to meet the needs for an accurate diagnosis, it is more difficult to biopsy again [11]. Thus, it is clinically important and necessary to explore a reliable, noninvasive, and practical method for the pre-therapy prediction of the histologic subtypes for treatment decision making and prognosis estimation in NSCLC patients.

Radiomics based on conventional medical images has been used to quantitatively assess tumor heterogeneity in more detail than visual analysis by analyzing the distribution and relationship of pixel or voxel gray levels in the lesion area [12, 13]. 18F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT)–based radiomics have been shown to have potential in differentiating ADC from SCC [14, 15]. Further studies have revealed that the discrimination performance could be further improved by combining with clinical features, like sex and smoking history (area under curve (AUC) = 0.859), which are higher than that of radiomic alone [16, 17]. However, only PET radiomic parameters were extracted and analyzed in the above studies. There is no single feature that can adequately describe the pathological phenotype of lesions due to the tumor heterogeneity [18].

Hence, the aim of this study was to develop and validate a prediction model, integrating the clinical characteristics, tumor marker levels [19], and radiomic features extracted from both the PET and CT images from the same volume of interest (VOI), for differentiating SCC from ADC in NSCLC and to provide a visually quantitative nomogram in clinical practice.

Materials and methods

Patients

We conducted a retrospective analysis of records from patients with NSCLC who were diagnosed by curative surgical resection between January 2017 and June 2019. This retrospective study was approved by the ethics committee of Shanghai Proton and Heavy Ion Center, and the requirement for informed consent was waived. The inclusion criteria included the following: (1) ADC or SCC that was confirmed by postoperative pathology according to the 2015 World Health Organization (WHO) classification [20], (2) standard routine whole-body PET/CT less than 30 days before surgery, and (3) single lesion with maximum standardized uptake value (SUVmax) > 2.50 and size > 1.00 cm. The exclusion criteria included the following: (1) previous history of malignant tumors and (2) anti-tumor therapy before PET/CT examination. We excluded 1385 patients among the 1700 patients with lung lesions initially recruited in our cancer center’s database to ensure the relationship between single pathological subtype and baselined clinico-biologico-radiological features. The patient recruitment process is presented in Fig. 1.

Fig. 1
figure 1

Flow chart showing the patient selection and exclusion

Full size image

Finally, totally 315 consecutive NSCLC patients were enrolled in this study, comprising 200 males and 115 females (mean age, 61.89 ± 9.10 years, range, 27–84 years), as summarized in Table 1. Patients were randomly divided into two independent sets according to a 7:3 ratio: training set (n = 220) and validation set (n = 95). Baseline data pertaining to clinical characteristics, including gender, age, height, weight, smoking status (never, ever/always), symptom (negative, positive chest symptoms), family history, the size, and location of lesion and the levels of serum tumor markers, including ferritin (FERR), squamous cell carcinoma antigen (SCCA), carbohydrate antigen 199 (CA 199), alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), cytokeratin 19 fragment antigen (CYFRA21-1), and neuron specific enolase (NSE) of each patient, were reviewed and recorded.

Table 1 Clinical and demographic characteristics of NSCLC patients
Full size table

18F-FDG PET/CT image acquisition and tumor segmentation

The 18F-FDG PET/CT scans were performed on a Biograph 16 PET/CT scanner (Siemens Healthcare, Erlangen, Germany) according to standard clinical scanning protocols. All patients fasted for at least 6 h before the scan, and none had a blood glucose levels > 8.7 mmol/L. A whole-body scan was acquired approximately 1 h after intravenous administering of 5.18 MBq/kg of 18F-FDG. The CT scans were performed first (120 kVp, 150 mAs, 0.33 s per rotation) using a slice thickness of 3.0 mm and reconstructed to a 512 × 512 matrix (voxel size: 0.98 × 0.98 × 3.0 mm3). Then, PET scans were performed with 2 min in each bed, a TrueX algorithm (2 iterations, 24 subsets, and 2 mm full width at half maximum) without filtering and smoothing was used to reconstruct the PET images. For all PET reconstructions, the matrix size was 200 × 200, resulting in anisotropic voxels of 4.07 × 4.07 × 3.0 mm3. The PET images were converted into SUV units by normalizing the activity concentration to the dosage of injected 18F-FDG and patient body weight.

Tumor segmentation was performed using Inveon Research Workplace (IRW, Siemens Healthcare, Erlangen, Germany) software. Two experienced nuclear medicine physicians drew boundaries in the axial, coronal, and sagittal PET scans that were large enough to include the primary tumor to delineate the VOIs using a threshold of 40% of SUVmax without knowing the pathology determined by consensus [21, 22]. To avoid the inclusion of areas with physiological 18F-FDG uptake within the regions of interest, a joint reading of both the CT and PET scans was performed side by side.

Quantitative radiomic feature extraction

The radiomic features were extracted using a voxel-based methodology. First, the SUV values contained within the VOIs were relatively resampled to 64 different values to yield a limited range of values, with the goal of reducing the noise and normalizing the images [23]. Then, totally 212 radiomic features were automatically calculated and extracted from the PET and CT images for each lesion using the Chang Gung Image Texture Analysis (CGITA) that is compliant to the Image Biomarkers Standardization Initiative, which is an open-source software code with a graphical user interface for radiomics running on MATLAB (version 2019a, MathWorks Inc., Natick, MA) (supplementary data Fig. S1) [24]. The details of radiomic features were described in supplementary data (Table S1).

Statistical analysis

The R (version 3.60, http://www.r-project.org) software was used for the statistical analysis. A comparison between the groups was performed using an independent t test or a Mann-Whitney U test for continuous variables and Fisher’s exact test or χ2 test for categorical variables. A two-sided p < 0.05 indicated statistical significance. Intra- and inter-class correlation coefficients (ICCs) were used to evaluate the consistency and reproducibility of the intra- and inter-observer agreements of the radiomic feature extractions. An ICC > 0.75 indicated good consistency.

Features selection and prediction model establishment

Univariate analysis was applied to identify the relevant predictors of the NSCLC subtypes in the training set. Multivariate analysis was performed by the least absolute shrinkage and selection operator (LASSO) binary logistic regression with 10-fold cross-validation, which was used to select the most useful factors [25, 26]. The prediction models that were performed to differentiate ADC from SCC were developed by the linear fusion of the selected non-zero features weighted by their coefficients, with prediction scores (Pre-scores) of each model calculated for each patient.

Prediction performance and clinical utility of prediction models

The performance of the models was evaluated by the receiver-operator characteristic curve (ROC) analysis and compared by the DeLong test. The AUC with 95% confidence interval (CI), sensitivity, specificity, and accuracy were calculated for each model. The clinical application value of the prediction models was determined and compared through the decision curve analysis (DCA) by quantifying the net benefit to the patient under different threshold probabilities in the queue.

Development and validation of individualized nomogram

To provide a visually quantitative tool to predict the histologic subtypes for NSCLC patients, we developed a nomogram on the basis of the prediction model with the highest AUC and clinical utility in the training set [27]. Calibration curves were plotted to assess the calibration of the nomogram by bootstrapping (1000 bootstrap resamples) based on the internal (training set) and external (validation set) validity.

Results

Clinical characteristics and tumor markers of patients

In total, 315 NSCLC patients comprising 122 SCC patients and 193 ADC patients were eventually enrolled in this study. The patients’ clinical characteristics and tumor markers of training set are summarized and compared in Table 2, while ones of validation set are provided in supplementary data (Table S2). SCC patients were more likely to be elderly males who had taller heights, a history of smoking, obvious symptoms, and larger lesions, while ADC patients were more likely to be younger females who had never smoked, no obvious symptoms, and smaller lesions (p < 0.05). The levels of FERR, SCCA, CYFRA21-1, and NSE in SCC patients were higher than those in ADC patients (p < 0.05). There were no significant differences in patient’s weight, family history, lesion location, and levels of CA199, AFP, and CEA between the SCC and ADC groups according to the univariate analysis (p > 0.05).

Table 2 Comparison of clinical characteristics and tumor markers between SCC and ADC patients in training set
Full size table

Features selection and prediction model establishment

A total of 315 regions with an increased 18F-FDG uptake were manually segmented, and 212 radiomic features were separately extracted by the two physicians. The agreement between the two physicians was excellent (all ICCs > 0.85, p < 0.05). Thus, the mean measurement values of the two physicians were used for further analysis.

For differentiating SCC from ADC, 4 independent prediction models (Clinical-Laboratory (Clin-Lab) Model, PET- Radiomic (Rad) Model, CT-Rad Model, and Combined Model) were built separately on the basis of selected clinical factors-tumor markers, PET radiomic parameters, CT radiomic parameters, and the combination of above features by LASSO regression in the training set (Fig. 2). The Pre-scores of each model for each patient were calculated using the following formulas:

$$ \mathrm{Pre}-\mathrm{score}\ \left(\mathrm{Clin}-\mathrm{Lab}\ \mathrm{Model}\right)=1.8145+0.8597\ast \mathrm{Sex}\ \left(\mathrm{Male}:0,\mathrm{Female}:1\right)-0.0847\ast \mathrm{Symptoms}\ \left(\mathrm{Negative}:0,\mathrm{Positive}:1\right)-0.3202\ast \mathrm{Size}\ \left(\mathrm{cm}\right)-0.0001\ast \mathrm{FERR}\ \left(\mathrm{ng}/\mathrm{mL}\right)-0.0020\ast \mathrm{SCCA}\ \left(\mathrm{ng}/\mathrm{mL}\right). $$
$$ \mathrm{Pre}-\mathrm{score}\ \left(\mathrm{PET}-\mathrm{Rad}\ \mathrm{Model}\right)=2.8790+1.4955\ast \mathrm{PET}\_{\mathrm{Coarseness}}^{\mathrm{Gray}\ \mathrm{Level}\ \mathrm{Neighborhood}\ \mathrm{Intensity}-\mathrm{difference}\ \mathrm{Matrix}\ \left(\mathrm{GLNIDM}\right)}+0.0025\ast \mathrm{PET}\_{\mathrm{Strength}}^{\mathrm{GLNIDM}}-0.0924\ast \mathrm{PET}\_\mathrm{Normalized}\_{\mathrm{Entropy}}^{\mathrm{Gray}\ \mathrm{Level}\ \mathrm{Co}-\mathrm{occurrence}\ \mathrm{Matrix}\ \left(\mathrm{GLCM}\right)}-0.4012\ast \mathrm{PET}\_{\mathrm{SUV}}_{\mathrm{min}}-0.1108\ast \mathrm{PET}\_{\mathrm{SUV}}_{\mathrm{mean}}-9.2039\ast \mathrm{PET}\_{\mathrm{Code}\ \mathrm{Similarity}}^{\mathrm{Texture}\ \mathrm{Feature}\ \mathrm{Coding}\ \mathrm{co}-\mathrm{occurrence}\ \mathrm{matrix}\ \left(\mathrm{TFCCM}\right)}+0.6994\ast \mathrm{PET}\_{\mathrm{Entropy}}^{\mathrm{Neighboring}\ \mathrm{Gray}\ \mathrm{Level}\ \mathrm{Dependence}\ \left(\mathrm{NGLD}\right)}. $$
$$ \mathrm{Pre}-\mathrm{score}\ \left(\mathrm{CT}-\mathrm{Rad}\ \mathrm{Model}\right)=1.7783\hbox{--} 0.0045\ast \mathrm{CT}\_\mathrm{Asphericity}-0.0024\ast \mathrm{CT}\_\mathrm{Entropy}\_\mathrm{prod}\_\mathrm{surface}\_\mathrm{area}+0.4938\ast \mathrm{CT}\_{\mathrm{Entropy}}^{\mathrm{NGLD}}. $$
$$ \mathrm{Pre}-\mathrm{score}\ \left(\mathrm{Combined}\ \mathrm{Model}\right)=3.6745+1.3360\ast \mathrm{Sex}\ \left(\mathrm{Male}:0,\mathrm{Female}:1\right)-0.2273\ast \mathrm{Size}\ \left(\mathrm{cm}\right)-0.0212\ast \mathrm{SCCA}\ \left(\mathrm{ng}/\mathrm{mL}\right)-0.0010\ast \mathrm{CYFRA}21.1\ \left(\mathrm{ng}/\mathrm{mL}\right)+0.0002\ast \mathrm{PET}\_{\mathrm{Strength}}^{\mathrm{GLNIDM}}+0.0001\ast \mathrm{PET}\_\mathrm{Normalized}\_{\mathrm{Contrast}}^{\mathrm{GLCM}}-0.0451\ast \mathrm{PET}\_\mathrm{Normalized}\_{\mathrm{Entropy}}^{\mathrm{GLCM}}-0.0498\ast \mathrm{PET}\_{\mathrm{SUV}}_{\mathrm{min}}-0.1138\ast \mathrm{PET}\_{\mathrm{SUV}}_{\mathrm{mean}}-0.0197\ast \mathrm{PET}\_\mathrm{Surface}\ {\mathrm{SUV}}_{\mathrm{mean}}\ 1+0.0469\ast \mathrm{PET}\_{\mathrm{Variance}}^{\mathrm{Texture}\ \mathrm{Feature}\ \mathrm{Coding}\ \left(\mathrm{TFC}\right)}-1.7315\ast \mathrm{CT}\_\mathrm{Second}\ {\mathrm{angular}\ \mathrm{moment}}^{\mathrm{TFCCM}}-0.4842\ast \mathrm{CT}\_{\mathrm{Correlation}}^{\mathrm{GLCM}}-0.0002\ast \mathrm{CT}\_\mathrm{Entropy}\_\mathrm{prod}\_\mathrm{surface}\_\mathrm{area}. $$
Fig. 2
figure 2

Features selection for prediction models using LASSO regression. The X-axis shows log (λ), and the Y-axis shows the model misclassification rate. The 5, 7, 3, and 14 features with non-zero coefficients are indicated for Clin-Lab Model (a), PET-Rad Model (b), CT-Rad Model (c), and Combined Model (d), respectively

Full size image

ADC patients generally had higher Pre-scores in all prediction models than those in SCC patients (p < 0.001) (Figs. 3 and 4). The selected radiomic features of the prediction models between ADC and SCC patients are summarized and compared in supplementary data (Table S3).

Fig. 3
figure 3

Violin plot of 4 prediction models for SCC and ADC patients in training set. The white dot represents the median. The black rectangle is the range from the lower quartile to the upper quartile. The black line running up and down through the violin diagram represents the range from the smallest non-outlier value to the largest non-outlier value

Full size image
Fig. 4
figure 4

Pre-scores of the Combined Model for each patient in training set

Full size image

Prediction performance and clinical utility of prediction models

The performance of these 4 prediction models to discriminate SCC from ADC is shown in Fig. 5. The Clin-Lab Model consisted of 3 clinical factors and 2 tumor markers, the PET-Rad Model consisted of 7 PET radiomic parameters, and the CT-Rad Model consisted of 3 CT radiomic parameters that were all significantly associated with the NSCLC pathological subtypes (AUCs (training set) = 0.887, 0.835, 0.784; AUCs (validation set) = 0.860, 0.740, 0.710, respectively).

Fig. 5
figure 5

Receiver-operating characteristic analysis of prediction models for predicting NSCLC subtypes in training set

Full size image

The DeLong test showed that the Combined Model, which consisted of 2 clinical factors, 2 tumor markers, 7 PET radiomic parameters, and 3 CT radiomic parameters, presented the optimal discrimination and best predictive sensitivity, specificity, and accuracy among the 4 models in both the training set (AUC (95% CI) = 0.932 (0.900–0.964), sensitivity = 96.25%, specificity = 95.00%, accuracy = 84.09%) and validation set (AUC (95% CI) = 0.901 (0.840–0.957), sensitivity = 93.55%, specificity = 81.25%, accuracy = 85.95%) (both p < 0.05) (Table 3).

Table 3 Performance of prediction models for predicting subtypes in NSCLC
Full size table

The DCA also showed that the Combined Model was the most reliable clinical treatment tool for predicting the histologic subtypes in NSCLC when the threshold probability was greater than 10% (Fig. 6).

Fig. 6
figure 6

Decision curve analysis (DCA) of prediction models in training set. The X-axis represented the threshold probability that was where the expected benefit of treatment was equal to the expected benefit of avoiding treatment. The Y-axis represented the net benefit. The gray and black line represented the hypothesis that all NSCLC patients were ADC and SCC, respectively

Full size image

Development and validation of individualized nomogram

According to the above results, we generated an individualized nomogram based on the Combined Model’s risk factors for the visualization (Fig. 7). The calibration curves of the nomogram for the probability of ADC demonstrated a good agreement between the prediction by the nomogram and the actual observation in both the training and validation sets (Fig. 8).

Fig. 7
figure 7

Developed the prediction nomogram based on Combined Model in training set. The probability of each predictor could be converted into scores according to the first scale “Points” at the top of the nomogram. After adding up the corresponding prediction probability at the bottom of the nomogram was the risk of ADC

Full size image
Fig. 8
figure 8

Calibration curves of nomogram in training (a) and validation (b) sets, respectively. The X-axis represented the predicted probability estimated by nomogram, whereas the Y-axis represented the actual observed rates. The solid line represented the ideal reference line that predicted NSCLC subtypes corresponds to the actual outcome, the short-dashed line represented the apparent prediction of nomogram, and the long-dashed line represented the ideal estimation. Calibration curves showed the actual probability corresponded closely to the prediction of nomogram

Full size image

Discussion

In this study, we successfully constructed and validated a Combined Model containing clinical factors, tumor markers, and radiomic features extracted from both the PET and CT images, which held an excellent performance in noninvasively stratifying NSCLC patients according to their pathological subtypes. In addition, we developed a visually quantitative nomogram for conveniently using this prediction model in clinical practice.

Of the clinical factors selected in the Combined Model, sex differences among NSCLC patients have been widely reported, with that SCC affecting more males than females [28]. The lesions are generally bigger in SCC patients than in ADC patients [29]. Tumor markers in serum are beneficial for the diagnosis and prognosis of NSCLC [30]. The serum levels of SCCA and CYFRA21-1 are highly sensitive in NSCLC and significantly higher in SCC than in ADC [31]. The results of this study are consistent with the conclusions of the above reports.

Different pathological subtypes lead to various clinical strategies and prognoses for NSCLC patients [5, 32]. The PET/CT-based radiomic is a relatively new quantitative imaging technique for the noninvasive assessment of tumors [33]. Ha S, et al. found that PET radiomic features were significantly different between ADC and SCC with 0.90 linear separability, but the study population was only 30 people [34]. Koyasu S, et al. also showed that PET radiomics was indeed useful in NSCLC subtypes with an AUC of 0.843 [15]. However, the radiomic approaches in the above studies were not be validated in another independent dataset. In this study, both the PET and CT radiomic approaches were applied and validated to have a good performance in the classification of NSCLC subtypes (AUCs (PET-Rad Model and CT-Rad Model) = 0.835 and 0.784, respectively). The above results indicated that the relationship between medical images and tumor molecular phenotypes can be established by radiomics, and then the diagnostic information of tumors can be obtained noninvasively through medical images for some patients who are not eligible for biopsy.

In addition, since radiomic extracts information from the tumor, an appropriate tumor segmentation algorithm is important for measuring tumor image parameters [35]. Ideally, the chosen segmentation method is both accurate and robust. Bashir et al. had compared various segmentation algorithms (freehand, 40% of maximum intensity threshold, and fuzzy locally adaptive Bayesian algorithms) in terms of inter-observer reproducibility and prognostic capability of texture parameters derived from NSCLC 18F-FDG PET/CT images [21]. They found that the models generated by all three segmentation algorithms were of at least equivalent utility. Moreover, segmentation with 40% of maximum threshold leads to the best reproducibility of image biomarkers when used by different observers. In this study, the agreements of the radiomic feature extraction using semiautomatic threshold-based methods were excellent (all ICCs > 0.85, p < 0.05). The high ability to reproduce and validate radiomic studies is vital to generating sufficient and convincing scientific evidence for translating potential applications into clinical practice [33, 36].

This study also explored whether the prediction performance based on radiomics could be further improved by combining with clinical factors and tumor marker levels. The Combined Model established in the present study not only significantly improved the prediction efficiency for subtype compared to these factors alone in both the training and validation sets (AUCs = 0.932 (training set), 0.901 (validation set), respectively) but also had higher performance than previous researches [14,15,16,17]. This discrepancy may be related to the complete and standard preoperative baseline data and postoperative pathological reports from a single center, as well as the appropriate algorithm [37]. The results of this study confirm the hypothesis and indicate that the heterogeneity of tumors can be evaluated more comprehensively by combining with multiscale characteristics of tumors, consistent with the report [38].

In addition, we generated an integrated nomogram on the basis of the Combined Model for facilitating its use in clinical practice. Clinical factors such as patient’s sex and age are recorded routinely at hospital admission. Moreover, we strongly recommend that serum tumor marker levels should be evaluated in patients who are highly suspected of having NSCLC or initially diagnosed with NSCLC, especially SCCA, CYFRA21-1. Both physicians and patients could perform a preoperative individualized prediction of the risk of ADC with this easy-to-use scoring tool, which can provide a noninvasive and accurate approach for patients who are unwilling or unable to undergo biopsy to develop more reasonable and effective treatment plans, especially the need of targeted therapy [39]. The DCA showed that if the threshold probability of a patient or doctor is > 10%, using this nomogram to predict the subtype adds more benefit than either the treat-all-patients as SCC or the treat-all-patients as ADC, which is more valuable for the current trend toward personalized medicine [40].

Although the results were encouraging, the present study had several limitations. Firstly, the sample selection was biased in this single-center retrospective study, and a new multicenter prospective study is still necessary to be designed for the further evaluation and verification of the generalization ability of the models. Secondly, some NSCLC patients, especially ADC patients, were excluded from the radiomic analysis due to the faint 18F-FDG uptake or small tumor size to ensure the quality of images and textural data. Small lesions are easier to be discovered in the early stage with the increasing use of imaging screening for lung cancer. Thus, a more sensitive tool that adaptively detects small tumors will be an important direction for future work. Finally, the patients with non-primary lung lesions were also excluded due to the purpose of this study. Noticeably that both primary and metastatic pulmonary nodules are very important for patients and clinical settings in the cancer center. The prediction model that widely used for lung lesions will be continually explored in future studies.

In conclusion, an integrated nomogram was constructed and validated in our study, which could provide a relatively accurate, convenient, and noninvasive method for the individualized discrimination between ADC and SCC in NSCLC patients, assisting in clinical decision making for precision treatment.