Prognostic value of 18F-FDG PET/CT radiomic model based on primary tumor in patients with non-small cell lung cancer: A large single-center cohort study

Objectives In the present study, we aimed to determine the prognostic value of the 18F-FDG PET/CT-based radiomics model when predicting progression-free survival (PFS) and overall survival (OS) in patients with non-small cell lung cancer (NSCLC). Methods A total of 368 NSCLC patients who underwent 18F-FDG PET/CT before treatment were randomly assigned to the training (n = 257) and validation (n = 111) cohorts. Radiomics signatures from PET and CT images were obtained using LIFEx software, and then clinical and complex models were constructed and validated by selecting optimal parameters based on PFS and OS to construct radiomics signatures. Results In the training cohort, the C-index of the clinical model for predicting PFS and OS in NSCLC patients was 0.748 and 0.834, respectively, and the AUC values were 0.758 and 0.846, respectively. The C-index of the complex model for predicting PFS and OS was 0.775 and 0.881, respectively, and the AUC values were 0.780 and 0.891, respectively. The C-index of the clinical model for predicting PFS and OS in the validation group was 0.729 and 0.832, respectively, and the AUC values were 0.776 and 0.850, respectively. The C-index of the complex model for predicting PFS and OS was 0.755 and 0.867, respectively, and the AUC values were 0.791 and 0.874, respectively. Moreover, decision curve analysis showed that the complex model had a higher net benefit than the clinical model. Conclusions 18F-FDG PET/CT radiomics before treatment could predict PFS and OS in NSCLC patients, and the predictive power was higher when combined with clinical factors.


Introduction
Lung cancer is one of the most predominant malignancies worldwide, with a high incidence and mortality rate. Non-small cell lung cancer (NSCLC) is the most common pathological type of lung cancer, accounting for 85% of all lung cancer cases (1). Lymph node metastases or distant metastases often occur before diagnosis due to the lack of early and specific clinical signs, significantly affecting treatment and prognosis. The 5-year survival rate for lung cancer is less than 20% (2), while the availability of new drugs or treatments has improved survival rates for lung cancer patients. Even with the same treatment, the patients reflect a vast difference. Identifying the risk of recurrence at the time of diagnosis allows the patient's treatment to be adjusted, thus enhancing the patient's survival.
Preoperative conventional imaging methods are essential for determining TNM staging, which is traditionally used to assess treatment response or survival. However, it has some limitations, which produce a significant variation in survival durations for patients with the same tumor stage. The prognosis of NSCLC remains uncertain, mainly at the time of diagnosis (3). Genomic studies have shown that intra-tumor heterogeneity is a common feature of solid tumors, such as NSCLC (4,5). However, because this procedure is invasive and sampling tissue is limited, it does not provide a complete characterization of tumor heterogeneity. Therefore, robust biomarkers that can provide a complete characterization of tumor heterogeneity may be of great value. Radiomics studies have shown promising results in revealing intra-tumor heterogeneity and predicting prognosis and treatment response in cancer patients (6)(7)(8). Radiomics is the high-throughput extraction of a large number of quantitative image features from medical images. It can capture information on the intensity, texture, and shape of lesions, which can be used to develop predictive models based on screened features to support personalized decision-making and treatment (9). 18 F-FDG PET/CT has been used by clinicians to diagnose and evaluate NSCLC patients (8). Several studies have focused on tumor radiomic analysis of NSCLC patients to predict recurrence (10)(11)(12). The combination of radiomic features (RSs) and TNM staging or other clinicopathology (CP) data performs better than TNM staging alone in estimating diseasefree survival in patients with early-stage NSCLC.
Recently, although radiomics has made significant progress in various malignancies (13)(14)(15), the radiomics literature for predicting prognosis in NSCLC is relatively sparse. A multimodality radiomics approach has great potential to address the limitations of unimodality models, as it can extract more meaningful features and thus provide a more robust description of the underlying tumor. In the present study, we retrospectively analyzed the data of NSCLC patients and screened feature parameters with predictive significance for prognosis by radiomics methods, aiming to construct a noninvasive deep learning model based on RSs of pretreatment PET/ CT images to predict the survival prognosis of NSCLC patients.

Patients
This retrospective study was approved by the institutional review board of the First Affiliated Hospital of Soochow University, and the informed consent was waived due to the retrospective nature of the study (ChiCTR2200062555).
NSCLC patients who underwent a combined imaging protocol of 18 F-FDG PET/CT between January 2019 and June 2021 were recruited in our present retrospective study. Inclusion criteria were as follows: (a) patients who underwent biopsy or surgery of lung tumor; (b) patients with immunohistochemistry (IHC) examination of PD-L1 performed; (c) histological type and grade were pathologically proven; (d) standard 18 F-FDG PET/CT was performed before biopsy or surgery; and (e) complete clinical profile. Exclusion criteria were as follows: (a) therapy (radiotherapy, chemotherapy, or chemoradiotherapy) was performed before 18 F-FDG PET/CT and IHC; (b) patients with unknown histological grade; (c) the size of the primary lesion was too small for segmentation; (d) multi-center primary lung cancer; and (e) patients with other types of cancers or with incomplete clinical and imaging datasets. A total of 368 patients were enrolled in the present study, and they were randomly divided into the training (n = 257) and validation (n = 111) cohorts at a ratio of 7:3 (16).

Acquisition and reconstruction of PET/ CT images
The 18 F-FDG PET/CT examination was performed after 6 h of fasting with blood glucose lower than 11.1 mmol/L. Approximately 40-60 min after the injection of 18 F-FDG (4.07-5.55 MBq/kg), PET/CT was performed from the base of the skull to the midthigh with 2-3 min per bed position (reconstructed by ordered subset expectation-maximization algorithm) using a Discovery PET/CT (General Electric Medical Systems, Milwaukee WI, USA) with low-dose CT parameters (140 kV, 120 mA, trans axial FOV of 70 cm, slice thickness 3.75 mm).

Feature extraction and selection
The region of interest (ROI) of the lesion was outlined layer by layer on the image cross-section using LIFEx freeware (17) (v6.30 https://www.lifexsoft.org/) two experienced nuclear medicine physicians who were blinded to the clinical and pathological information of patients. Areas with abnormal uptake of 18 F-FDG on PET and abnormal density on CT were defined as lesions. The volume of interest (VOI) in threedimensional coordinates was automatically defined on PET images with the SUV max threshold of 41%. The RSs were summarized across all VOIs. Spatial resampling had a voxel size of 2 × 2 × 2 mm. Intensity discretization for CT data was performed with the number of gray levels of 400 bins and absolute scale bounds from -1,000 and 3,000 HU, while it was conducted with 64 bins between 0 and 20 for PET data. The RSs were extracted from PET and CT images within the same VOI due to the excellent matching of PET and CT images. If the respiratory motion caused a mismatch between the PET and CT images, it could be adjusted manually. The RSs of PET and CT images were selected by the following procedure. Intra-class correlation coefficients (ICCs) were calculated to evaluate the repeatability of all the features, of which ICCs > 0.75 were selected for model construction (18,19). Subsequently, the retained features were further selected by the least absolute shrinkage and selection operator (LASSO) regression algorithm. In addition, 10-fold cross-validation was applied to select the parameter of Lambda (l) to avoid overfitting.

Patient follow-up
Patients were followed up from the first diagnosis of NSCLC by surgery or puncture pathology to June 30, 2021 (the last follow-up date), in conjunction with outpatient review data or by telephone. Progression-free survival (PFS) was defined as the time from the first diagnosis of NSCLC by surgery or puncture pathology to the date of tumor progression, death, or the followup cut-off date. Overall survival (OS) was defined as the time from the first diagnosis of NSCLC by surgery or puncture pathology to the date of death or the follow-up cut-off date. Receiver operating characteristic (ROC) curves and Harrell's consistency index (C-index) were used to assess the model's capability to predict outcomes in both training and validation cohorts. Moreover, calibration curves were produced to assess the model's calibration. In order to evaluate the clinical value of various models, decision curve analysis (DCA) was employed to calculate the net benefit at various threshold probabilities.

Statistical analysis
IBM SPSS statistics version 26.0, Python version 3.0 (https:// www.python.org), MedCalc software (MedCalc Software, Ostend, Belgium), and R 3.6.3 software package were used for statistical analyses. P < 0.05 was considered statistically significant. The continuous variables were transformed into dichotomous variables for further statistical analysis. The Kaplan•Meier (KM) curve was used for survival analysis, the log•rank test was used for univariate survival analysis, Cox proportional risk regression model was used for the multifactor prognostic analysis of significant influencing factors, and the hazard ratio (HR) and its 95% CI were estimated.

Patient characteristics
A total of 368 patients were enrolled in this study.

Cox analysis of prognostic influences
Univariate and multivariate Cox regression analyses were used to determine the relationship between clinical features and imaging indicators with survival ( Table 2 and Table 3; Figure 2). Using PFS as the study endpoint, KM univariate analysis showed that gender, TNM staging, pathology, smoking history, treatment modality, distant metastasis, Ki-67, lymph node metastasis, SUV max , MTV, TLG, and Rad-score were factors influencing the PFS of patients (all P < 0.05). Cox multifactorial analysis showed that gender, smoking history, lymph node metastasis, distant metastasis, and Rad-score were independent prognostic factors affecting PFS (all P < 0.05). With OS as the study endpoint, KM univariate analysis showed that gender, PD-L1 50%, smoking history, distant metastasis, lymph node metastasis, SUV max , MTV, TLG, and Rad-score were factors influencing the OS of patients (all P < 0.05). Cox multivariate analysis showed that smoking history, distant metastasis, and Rad-score were factors influencing the OS of patients (all P < 0.05). Furthermore, Rad-score and clinical factors were significantly correlated with PFS and OS in the training cohort and validation cohort ( Figure 3 and Figure 4).

Assessment and validation of the models for predicting PFS and OS
To provide a quantitative method to predict patients' probability of 1-, 2-, and 3-year PFS and OS, a radiomics nomogram combining clinical factors was established ( Figure 5).
For the training cohort, the C-index and AUC of the C-R model were 0.775 (95% CI: 0.722-0.828) and 0.780 (95% CI: 0.724-0.829) for  KM analyses for PFS (A-E) and OS (F-H) according to gender, metastasis, smoking history, R-signatures, and lymph node metastasis in the training cohort. Clinical use Figure 6 presents the DCA for the radiomics model and clinical model. DCA indicated that the net benefit of the radiomics nomogram was faintly higher compared with the clinical nomogram for both PFS and OS.

Discussion
Currently, lung cancer remains the leading cause of cancerrelated deaths worldwide (1). The survival of NSCLC patients varies greatly among individuals, and the survival prognosis is affected by many factors. Our study found that Rad-scores on Traditional prognostic assessment methods, such as TNM staging, are important in guiding the formulation of rational treatment plans and prognosis assessments for lung cancer patients (20)(21)(22). However, even at the same TNM stage, the individual differences between patients make the prognosis significantly different, and the differences are magnified as the stage increases. Therefore, it is difficult to accurately assess the survival of NSCLC patients based on the TNM staging system alone. 18 F-FDG PET/CT is a whole-body imaging examination that is widely used for the staging of malignant tumors. Previous studies have found that tumor FDG uptake can be used as a prognostic marker in patients with advanced and inoperable early-stage NSCLC (23, 24). Meta-analysis of IASLC shows that patients with higher preoperative primary tumor SUV max have lower survival rates and shorter survival times (25). Konings R et al. have shown that the primary tumor SUV max has a predictive value for the prognosis of NSCLC (26,27). As the  SUV-based approach is influenced by many physiological and technical factors, such as patient preparation, coordination of image acquisition, reconstruction, and analysis (28,29), other features derived from PET/CT images are currently being explored. Radiomics is the comprehensive quantification of tumor phenotypes through the application of numerous quantitative imaging features, which may reflect changes in human tissues at the cellular and genetic levels and offer more thorough details on tumor biology and the microenvironment in addition to visual features (30,31). Radiomics has been used for the diagnosis, response assessment, and survival prognosis of various cancers (32)(33)(34). Currently, there are few studies on PET/CT imaging histology to predict the prognosis of NSCLC. It has been reported that imaging features are an independent prognostic factor for NSCLC and are associated with multiple clinical endpoints (35,36). Therefore, we attempted to establish a complex model to assess the potential prognostic value of NSCLC patients by combining PET/CT-based radiomics features with clinical features.
Our study investigated the potential prognostic value of imaging histological features derived from pretreatment 18 F-FDG PET/CT images in NSCLC patients. In this study, a complex model combining radiomics and clinical factors was found to be more effective in predicting PFS and OS, improving the efficacy of traditional clinical parameters in assessing the prognosis of NSCLC and aiding the prognosis of NSCLC patients. These findings were consistent with previous studies showing that RSs extracted from PET/CT images have potential implications for assessing disease status and risk stratification in NSCLC (37, 38).
The present study has a number of limitations. Firstly, although the patients included in this study were the largest patient group compared with previous studies, the sample size was relatively small, and it was a retrospective study. Secondly, the model was developed and validated based on a single institution dataset. The relationship between clinicopathological parameters and prognosis might have been influenced by the sample size. Future prospective multi-center studies with large data are required to obtain more accurate and reproducible imaging histological features. Finally, the effectiveness of adjuvant therapy for patients with high Rad-scores according to 18 F-FDG PET/CT nomograms should also be evaluated in the future.

Conclusions
In conclusion, a complex model combining clinicopathological factors and radiomic features extracted from 18 F-FDG PET/CT images had the potential to predict PFS and OS, and the nomogram analyses could provide more accurate individualized predictions of PFS and OS for patients, which in turn had the potential to help clinicians make more informed decisions in clinical practice.

Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Ethics statement
This retrospective study was approved by the institutional review board of the First Affiliated Hospital of Soochow University, and the informed consent was waived due to the retrospective nature of the study (ChiCTR2200062555). The ethics committee waived the requirement of written informed consent for participation. B A FIGURE 6 DCA of PFS (A) and OS (B) for the model. The y-axis measures the net benefit. The net benefit was calculated by summing the benefits (true positive results) and subtracting the harms (false-positive results), weighting the latter by a factor related to the relative harm of undetected cancer compared with the harm of unnecessary treatment. The C-R model had the highest net benefit compared with the clinical model.

Author contributions
JL: data curation and writing-original draft preparation. BZ: conceptualization, methodology, and software. SG: visualization and investigation. SD: supervision. CH: software and validation. SS: writing-reviewing and editing. All authors contributed to the article and approved the submitted version.