The predictive value of [¹⁸F]FDG PET/CT radiomics combined with clinical features for EGFR mutation status in different clinical staging of lung adenocarcinoma

We retrospectively and consecutively collected lung cancer patients who underwent [¹⁸F]FDG PET/CT examinations before treatment in our hospital from January 2018 to April 2022. Inclusion criteria: (1) Lung adenocarcinoma was confirmed by surgery or biopsy pathology, and the pathological classification was based on IASLC/ATS/ERS lung adenocarcinoma classification criteria [36]; (2) the patients completed [¹⁸F]FDG PET/CT examination before surgery, and the interval between surgery and examination was less than 30 days; (3) the EGFR mutation test result was available; (4) patient had no history of other malignant tumors. Exclusion criteria: (1) lesions with poor image quality or difficulty to measure; (2) patient had other subtypes of lung cancer; (3) no chest CT images.

According to the above criteria, five hundred fifteen patients with lung adenocarcinoma were included, and the clinical information of age, gender, smoking status, clinical stage, tumor marker, SUV_max, and postoperative pathology was recorded. We used patients collected from January 2018 to April 2021 as the training set and patients collected from May 2021 to April 2022 as the testing set. The study was performed in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. The study protocol was approved by the Ethics Committee of our institution (No. [2022] KD 087). As our study is a retrospective analysis, informed consent is not required. The study flow chart is shown in Fig. 1.

EGFR mutation test was performed on tissue specimens obtained by surgical resection or biopsy. Real-time fluorescent PCR was performed to detect EGFR mutations in exons 18–21 using the EGFR gene mutation detection kit purchased from Shanghai Yuanqi Company. The detailed procedures followed the manufacturer's instructions (see Additional file 1). If the mutation was detected in any of the above exons, the lesion was defined as EGFR mutant; otherwise, the lesion was defined as EGFR wild type.

The image acquisition protocol was developed according to the acquisition protocol based on Imaging Biomarker Standardization Initiative (IBSI) guidelines [37]. The PET/CT image acquisition instrument was a German Siemens BiographmCT (64) PET/CT machine. The patients fasted for 4–6 h before the examination, and their height, weight, and blood sugar were recorded on the examination day. [¹⁸F]FDG was intravenously injected according to the patient's body weight at 3.70–5.55 MBq/kg. The imaging agent was provided by Nanjing Jiangyuan Andico Positron Research and Development Company (radiochemical purity > 95%). The patients underwent PET/CT whole-body imaging after resting in a quiet and comfortable environment for 1 h. The patient was placed in the supine position and kept holding the head with both hands. Imaging lasted for 2 min/bed, and the collection range was from the base of the skull to the middle of the femur. Diagnostic chest CT imaging was performed after the PET/CT scan. After image acquisition, the TrueX + TOF (ultraHD-PET) system was used for image reconstruction. A postprocessing workstation TrueD system (Siemens) was used for image evaluation. Image acquisition parameters are listed in Additional file.

[¹⁸F]FDG PET/CT image analysis: PET and CT images were analyzed by two physicians (A and B) with 3 years of experience in nuclear medicine. The software they used was the 3D slicer software 4.11.2 (http://​www.​slicer.​org). For PET images, they used a semi-automatic segmentation method developed by Beichel et al. [38]. For CT images (3 mm), they used NVIDIA AI-Assisted Annotation (3D-Slicer built-in) and the boundary-based CT segmentation models to process lung nodule images. The segmented tumor region of interest (ROI) was checked and proofread by another physician with more than 10 years of experience in PET/CT diagnosis. Four weeks after completing the ROI for all cases, Doctor A segmented the tumor region again for 300 patients, among which 70 patients were randomly selected for Doctor B to segment.

Before feature extraction, the image was normalized, and all images were interpolated (sitkBSpline algorithm, 3rd-order B-spline interpolation) so that the isotropic voxel spacing was rotated unchanged and the extracted features were compared between different samples. CT images were resampled to 1 × 1 × 1 mm³, and PET images were resampled to 3 × 3 × 3 mm³. The method of fixed bin width was used for discretization, and the bin width of CT and PET images was 25 and 0.313, respectively. The bin discretization, Laplacian of Gaussian (LOG), and wavelet transform were applied to generate different feature sets. For the LOG filter, different sigma values were used to extract fine, medium, and coarse features; specifically, these features ranged from 0.5 to 5 with a step size of 0.5. The wavelet transforms produced 8 decompositions per layer (applying all possible combinations of high-pass or low-pass filters in each of the three dimensions, including HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL). All intensity, histogram, and texture features were preprocessed (including discretization, logarithm, and wavelet).

Using the Pyradiomics module in Python 3.8.8, we extracted multiple features from different feature categories based on three segmented ROIs (twice from Doctor A and once from Doctor B). These categories included shape and morphology-based features (14 shape features), first-order statistics (18 FOS features), gray-level co-occurrence matrix (GLCM 24 features), gray-level dependency matrix (14 GLDM features), gray-level run-length matrix (16 GLRLM features), gray-level size region matrix (16 GLSZM features), and the neighbor gray-level tone difference matrix (5 NGTDM features). The features extracted from three sets of ROIs were assessed for within-group and between-group intraclass correlation efficient (ICC), and the features with ICCs greater than 0.75 were considered in good consistency and kept for further analysis.

To avoid overfitting, the variance method was used to remove features with small variance (threshold = 0.24). Next, the Mann–Whitney U test was used in the training set to screen out radiomics features with p value < 0.1, which might be associated with EGFR mutation status. Then, the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was applied to the normalized training set data to select the best predictive features. The LASSO algorithm added an L1 regularization term to the least squares algorithm to avoid overfitting and employed fivefold cross-validation.

Machine learning models were built using the Sklearn module in Python 3.8.8. Nine models were constructed based on CT, PET, and PET/CT radiomics features with LR, SVM, and RF, respectively. The training set used a grid search with fivefold cross-validation to find the optimal hyperparameters (the specific parameters are listed in Additional file 1: Table S1), and the models were retrained on the entire training set. Receiver operating characteristic (ROC) curves and the area under the curve (AUC) were used to evaluate model performance in training and testing sets. The 3 models with the best performance on the testing set were kept to calculate the Rad-score. We used the SHAP module in Python 3.8.8 to interpret the model better to understand the importance of different features in different models. The biggest advantage of the SHAP value is that SHAP can reflect the positive and negative influence of the features in each sample. The radiomics workflow used in this study is shown in Fig. 2.

Statistical analysis was performed using R software (version 3.4.3; http://​www.​R-project.​org/​). Continuous variables were expressed as mean ± standard deviation (normal distribution) or median (Q1–Q3) (skewed distribution). Categorical variables were expressed as frequency or rate (%). Differences in clinical data and PET/CT metabolic parameters between different EGFR mutation status (binary variables) were tested using the χ² test (for categorical variables), T test (for normal distribution), or Mann–Whitney U test (skewed distribution). Multivariate logistic regression method was used to construct a clinical model with significant clinical parameters, and a joint model and the corresponding nomogram were built by combining the clinical parameters with three Rad-scores. The minimum Akaike information criterion was used to select the best model parameters. The calibration curve confirmed the agreement between the nomogram and observations, and the model’s power was evaluated using ROC curve and AUC. The model's accuracy, sensitivity, and specificity were calculated to obtain a quantitative performance measurement. Decision curve analysis (DCA) assessed the models' clinical utility and net clinical benefit. A pairwise comparison of the model AUCs was performed using the method proposed by Delong et al. [39]. All statistical tests were two-sided, and p < 0.05 was considered statistically significant. Since the carcinoembryonic antigen (CEA) of 49 cases (9.5%), the cytokeratin 19 fragment (CYFRA21-1) of 99 cases (19.1%), the neuron-specific enolase (NSE) of 74 cases (14.3%), the serum squamous cell carcinoma antigen (SCC-Ag) of 91 cases (17.6%) were missing, we imputed the missing data using miceforest (version 5.4.0; https://​github.​com/​AnotherSamWilson​/​miceforest/​).

The clinical characteristics of the patients are shown in Table 1. A total of 515 patients with lung adenocarcinoma were enrolled in the study, including 264 females and 251 males. The average age was 64.0 ± 9.2 years (ranging from 36 to 87 years). In total, 175 patients (34.0%) had a smoking history, and 348 (67.6%) had solid nodules. The patients’ clinical stages were: stage I: 209 cases (40.6%), stage II: 24 cases (4.7%), stage III: 85 cases (16.5%), and stage IV: 197 cases (38.3%). The EGFR mutation status was pathologically confirmed by surgical resection or biopsy: there were 202 cases (39.2%) with EGFR wild type and 313 cases (60.8%) with EGFR mutant (including 3 cases of exon 18, 127 cases of exon 19, 13 cases of exon 20, 150 cases of exon 21, 1 case of exon 19 + 20, 2 cases of exon 19 + 21, 1 case of exon 20 + 21, and 16 cases of the unknown exon).

There were no significant differences between the training set (n = 404) and testing set (n = 111) in age, gender, smoking history, nodule type, nodule location, tumor markers, and EGFR mutation rate (all p > 0.05). Tumor long axis, tumor short axis, clinical stage, and SUV_max showed significant differences between the two datasets (all p < 0.05), possibly due to the different compositions of patients during different periods (see Additional file 1: Table S2). To eliminate this difference, we performed a stratified analysis (stratified by clinical stage) in both datasets to verify the robustness of the joint model. The clinical parameters of gender, smoking history, nodule type, CEA, SCC-Ag, clinical stage, tumor long axis, and tumor short axis explained the differences between EGFR mutant and wild-type patients (all p < 0.05 in training set).

SUV_max was significantly different between EGFR mutant and wild type in the training set (p = 0.005) but not in the testing set (p > 0.05). SUV_max had a weak predictive ability for EGFR mutation status in lung adenocarcinoma (training set AUC = 0.582, 95% CI 0.526–0.638, testing set AUC = 0.584, 95% C0I 0.475–0.694).

Based on the segmentations of tumor regions on PET/CT images, a total of 3562 radiomics features (1781 PET features, 1781 CT features) were extracted. Among them, 423 CT features and 248 PET features were excluded based on intragroup ICCs evaluation, and 109 CT features and 81 PET features were excluded according to intergroup ICCs evaluation. Next, we used the variance method, Mann–Whitney U test, and LASSO algorithm (Additional file 1:Figure S1) to further screen out 8 CT features, 4 PET features, and 4 PET/CT fusion radiomics features (2CT + 2PET), respectively (Additional file 1:Table S3).

The corresponding AUCs of the radiomics models are shown in Fig. 3, and the feature weights for each model are shown in Additional file 1:Figure S2. Among the 9 constructed radiomics models, the three models based on RF algorithm were better than the models based on LR and SVM algorithms in both training set and testing set, and the AUCs of CT_RF, PET_RF, and PET/CT_RF in the training set were 0.688, 0.666, and 0.698, respectively, and the AUCs in the testing set were 0.726, 0.678, and 0.704, respectively. Although the performance of CT_RF model was better than that of PET_RF and PET/CT_RF models in the testing set, the difference was not significant (both p > 0.05). Table 2 lists the diagnostic efficacy of the three best models.

We further compared the performance of CT_RF, PET_RF, PET/CT_RF, and SUV_max for predicting EGFR mutation status. The ROC curves of the three radiomics models and SUV_max in training set and testing set are shown in Additional file 1: Figure S3. The AUCs of three radiomics models were significantly better than SUV_max in both training and testing sets (all p < 0.05).

We first constructed a clinical prediction model (baseline model) with the 8 clinical parameters in Table 1 that might be related to the EGFR mutation status. Five parameters were finally included, and the model equation is as follows:

Logit(P) = 0.91617 + 0.00156 × CEA − 0.05743 × SCC-Ag − 0.92507 × (Gender = male) − 0.57848 × (Smoking history = positive) + 0.70786 × (Nodule Type = sub-solidity).

Next, the above 5 parameters were combined with the three best radiomics models (CT_RF, PET_RF, and PET/CT_RF) to construct joint prediction models. The three joint models are as follows:

CT joint model:

Logit(P) = − 2.11428 + 0.00177 × CEA − 1.31985 × (Gender = male) + 0.52447 × (Nodule Type = sub-solidity) + 4.96787 × CT_Rad;

PET joint model:

Logit(P) = − 1.01180 + 0.00170 × CEA –0.86919 × (Gender = male)− 0.58352 × (Smoking history = positive) + 4.02651 × PET_Rad;

PET/CT joint model:

Logit(P) = − 1.66386 + 0.00183 × CEA –0.89610 × (Gender = male) − 0.53267 × (Smoking.history = positive) + 5.20912 × PET/CT_Rad.

The ROC and DCA curves of the clinical and three joint models are shown in Fig. 4, and the diagnostic efficacy is shown in Table 2. In the training set, CT_joint model had the highest specificity of 0.795, while the PET joint model and the PET/CT joint model had the highest accuracy of 0.713, and the PET joint model had the highest sensitivity of 0.794. In the testing set, CT_RF had the highest specificity of 0.756, while the PET/CT_RF and the PET/CT joint model had the highest accuracy of 0.712, and the PET/CT_RF model had the highest sensitivity of 0.800.

The AUCs of three joint models and clinical model in the training set followed the order of CT joint model > PET/CT joint model > PET joint model > clinical model, but only the AUC of CT joint model was significantly better than the clinical model (p = 0.049). The AUC values in the testing set followed the order of PET/CT joint model > CT joint model > PET joint model > clinical model, but the differences were not significant (p > 0.05). By calculating the net reclassification index (NRI), we found that the PET/CT joint model correctly reclassified 10.0% (95% CI 1.0–19.0%, p = 0.029) of mutant type more than the clinical model, while there was no significant difference in identifying wild type (p = 0.410). Further comparison of DCA showed that the net benefit of the three joint models was higher than that of the clinical model in both training and testing sets (Fig. 4).

We further performed stratified analysis to verify the diagnostic performance of radiomics models, clinical models, and joint models in different clinical stage (Table 3). In the training set, clinical stage significantly affected the prediction of EGFR mutation status by the three radiomics models and the joint PET/CT model, suggesting there was an interaction effect (all p < 0.05); however, this effect was not significant in the testing set (p = 0.067–0.869). The EGFR mutation rates and clinical characteristics of the patients with different clinical stages are shown in Additional file 1: Table S4. For stage I-II nodules, besides clinical model, other models all showed varying degrees of overfitting. The CT joint model performed best in the training set (AUC: 0.838), while the CT_RF model performed the best in the testing set (AUC: 0.797). For stage III-IV nodules, the clinical model performed the best in the training set (AUC: 0.729), followed by the PET/CT joint model (AUC: 0.722); in the testing set, the clinical model showed significant overfitting (AUC: 0.675), while the PET/CT joint model showed the best performance (AUC: 0.723). The nomogram and a calibration curve of the PET/CT joint model are shown in Fig. 5.