Development of lung cancer risk prediction models based on F-18 FDG PET images

We aimed to evaluate whether the degree of F-18 fluorodeoxyglucose (FDG) uptake in the lungs is associated with an increased risk of lung cancer and to develop lung cancer risk prediction models using metabolic parameters on F-18 FDG positron emission tomography (PET). We retrospectively included 795 healthy individuals who underwent F-18 FDG PET/CT scans for a health check-up. Individuals who developed lung cancer within 5 years of the PET/CT scan were classified into the lung cancer group (n = 136); those who did not were classified into the control group (n = 659). The healthy individuals were then randomly assigned to either the training (n = 585) or validation sets (n = 210). Clinical factors including age, sex, body mass index (BMI), and smoking history were collected. The standardized uptake value ratio (SUVR) and metabolic heterogeneity (MH) index were obtained for the bilateral lungs. Logistic regression models including clinical factors, SUVR, and MH index were generated to quantify the probability of lung cancer development using a training set. The prediction models were validated using a validation set. The lung SUVR and lung MH index in the lung cancer group were significantly higher than in the control group (p < 0.001 and p < 0.001, respectively). In the combined prediction model 1, age, sex, BMI, smoking history, and lung SUVR were significantly associated with lung cancer development (age: OR 1.07, p < 0.001; male: OR 2.08, p = 0.015; BMI: OR 0.93, p = 0.057; current or past smoker: OR 5.60, p < 0.001; lung SUVR: OR 1.13, p < 0.001). In the combined prediction model 2, age, sex, BMI, smoking history, and lung MH index showed a significant association with lung cancer development (age: OR 1.06, p < 0.001; male: OR 1.87, p = 0.045; BMI: OR 0.93, p = 0.010; current or past smoker: OR 4.78, p < 0.001; lung MH index: OR 1.33, p < 0.001). In the validation data, combined prediction models 1 and 2 exhibited very good discrimination [area under the receiver operator curve (AUC): 0.867 and 0.901, respectively]. The metabolic parameters on F-18 FDG PET are related to an increased risk of lung cancer. Metabolic parameters can be used as biomarkers to provide information independent of the clinical parameters, related to lung cancer risk.


Introduction
Lung cancer is the leading cause of cancer-related death worldwide [1].According to the National Lung Screening Trial (NLST) results, medical organizations have recommended annual CT lung cancer screening in high-risk individuals aged 55-74 years who have a 30 pack-year smoking history and currently smoke or have quit within 15 years [2].However, lowdose CT screening reveals a large number of indeterminate nodules, and < 50% of patients with lung cancer are eligible for screening [3].In addition, major concerns regarding annual low-dose CT screening include additional medical costs due to a large number of false-positive results and the harm caused by excessive radiation exposure from repeated CT scans [4].Thus, selective low-dose CT screening in individuals with a higher risk of lung cancer is expected to improve the costeffectiveness and reduce harm.
In the pathophysiology of lung cancer, it is now evident that chronic inflammation is involved in all stages of lung cancer development, from malignant transformation and tumor initiation to invasion and metastasis [5].Persistent exposure to several exogenous factors, including smoking, asbestos, radiation, and air pollution, increases the risk of chronic inflammation and consequently induces lung cancer development in combination with genetic factors [6].From the viewpoint of lung cancer pathophysiology with chronic pulmonary inflammation, F-18 fluorodeoxyglucose (FDG) positron emission tomography (PET) can provide additional information to identify individuals with a higher risk of lung cancer because it can be used to noninvasively evaluate the inflammatory activity in the body [7].In particular, F-18 FDG PET can serve as a surrogate marker of pulmonary inflammation [8].
Thus, it is postulated that the specific pattern of F-18 FDG uptake in the lungs, indicating chronic pulmonary inflammation, can be used to predict lung cancer development.The aim of the present study was to evaluate whether the degree of F-18 FDG uptake in the lungs is associated with an increased risk of lung cancer, and to develop lung cancer risk prediction models using metabolic parameters on F-18 FDG PET.

Study population
This retrospective study included healthy individuals who underwent initial F-18 FDG PET/CT scans for a health check-up at our institution between January 2008 and December 2016; we confirmed no evidence of lung cancer or other malignancies.Individuals who met the following criteria were excluded: no follow-up chest CT or PET/CT after 5 years of the initial PET/CT; no medical record of smoking history or body mass index (BMI); and history of acute infection.Individuals who developed lung cancer within 5 years of the initial PET/CT scan were classified into the lung cancer group.The development of lung cancer was defined based on the follow-up chest CT or PET/CT scans, and confirmed by pathological results.Individuals who did not develop lung cancer within 5 years after the initial PET/ CT scan were classified into the control group.Clinical factors including age, sex, BMI, and smoking history (never, past, or current smoker) were collected for all patients.The healthy individuals were then randomly assigned to either the training or validation sets.This study was approved by the Institutional Review Board of our institution.

F-18 FDG PET/CT images
A PET/CT system (Discovery STE 16, GE Healthcare, Milwaukee, WI, USA) was used to acquire torso F-18 FDG PET images, as previously described [9].Patients received intravenous administration of 7.0 MBq/kg F-18 FDG.Images were acquired 60 min after F-18 FDG administration.A noncontrast CT scan was performed for attenuation correction and localization.Immediately after the CT scan, PET images were acquired from the base of the skull or the top of the brain to the proximal thigh.
To develop the risk prediction model, an experienced nuclear medicine physician (H.W.K.) reviewed all F-18 FDG PET images using a dedicated workstation (Advantage Workstation version 4.3; GE Healthcare).Two-dimensional regions of interest (ROIs) (long-axis diameter range, 10 mm) were drawn on the pulmonary parenchyma in the left upper lobe (LUL), left lower lobe (LLL), right upper lobe (RUL), right middle lobe (RML), and right lower lobe (RLL) of the lung, as well as the blood pool of the aortic arch (Fig. 1).The ROIs included chronic inflammatory changes such as emphysematous changes, while excluding pulmonary nodules, vascular structures, active inflammation, and bullae.For each lobe of the bilateral lungs, the maximum standardized uptake value (SUVmax), mean SUV (SUVmean), and minimum SUV (SUVmin) were obtained.The SUV ratio (SUVR) and metabolic heterogeneity (MH) index were calculated as follows: SUVR = (SUVmax of lung parenchyma/ SUVmax of blood pool) × 100; MH index = [(SUVmax of lung parenchyma − SUVmin of lung parenchyma)/SUVmax of blood pool)] × 100, according to a previously validated method [10].The lung SUVmax, SUVmean, SUVmin, SUVR, and MH index were defined as the respective average values for each lobe.In addition, an ROI (long-axis diameter range, 10 mm) was drawn on the central area of the right hepatic lobe, to obtain the liver SUVmax, SUVmean, and SUVmin.To assess inter-observer reproducibility, one another experienced nuclear medicine physician (K.S.W.) reviewed the F-18 FDG PET images of 100 randomly selected participants in the training set.The ROIs of the bilateral lungs were delineated using the same methodology as in the development of the risk prediction model, and the lung SUVmax, SUVmean, SUVmin, SUVR, and MH index were calculated accordingly.

Development of the risk prediction model
The prediction models were configured to estimate the absolute risk that an individual would have lung cancer in 5 years.Logistic regression models to quantify the probability of lung cancer development were generated using the training data set.Simple prediction models were constructed, including only clinical factors (age, sex, BMI, and smoking history) or only metabolic parameters (lung SUVR and lung MH index).Combined prediction models, including clinical factors and metabolic parameters, were also constructed.The combined prediction model 1 included all variables in the clinical model and lung SUVR.In the combined prediction model 2, all variables in the clinical model and the lung MH index were included.The β coefficients from the logistic regression models were used to create the prediction model: probability = e k /(1 + e k ), where k = β 0 + β 1 X 1 + β 2 X 2 + ⋯ + β n X n .

Validation of the risk prediction model
The prediction models were validated using the validation data set.The performance of the prediction models was evaluated using C-statistics with respect to discrimination.The C-statistic, also called the receiver operating characteristic (ROC) curve, is a measure of goodness of fit for binary outcomes in a logistic regression model [11].The prediction model was considered good when the area under the ROC curve (AUC) was over 0.70 and very good when the AUC was over 0.80 [12].The performance of the prediction models was also evaluated using the Brier score with respect to calibration.Calibration numerically represents how closely the predicted probabilities match the actual outcomes.The risk of developing lung cancer for each participant was calculated, and the average predicted probabilities were compared with the actual probabilities of developing lung cancer in each decile.The Brier score was calculated as the mean squared difference between the predicted and actual outcome.A Brier score of 0 indicates perfect accuracy.

Statistical analysis
All statistical analyses were conducted using SPSS for Windows, version 25.0 (IBM Corp., Armonk, NY, USA).Pearson's correlation analysis was employed to assess the associations between metabolic parameters, including SUVmax, SUVmean, SUVmin, SUVR, and MH index, with age or BMI.Two-sample t-tests were utilized to compare all metabolic parameters for the lung, liver, and blood pool, between male and female groups, smoker and non-smoker groups, as well as control and lung cancer groups.Paired sample t-tests were performed to compare metabolic parameters among each lobe of the bilateral lungs.Univariate and multivariate logistic regression models, incorporating both metabolic parameters and clinical factors, were utilized for the development of the risk prediction model.Inter-observer reproducibility was assessed using the coefficient of variation (CV) between the two observers.A CV of 15% or lower was considered indicative of good reproducibility [13].Variables with p-values < 0.05 from the univariate analysis were included in the subsequent multivariate regression analysis.A p-value < 0.05 was considered statistically significant.

Characteristics
A total of 795 participants from our institution were enrolled in this study.Of these, 136 developed lung cancer within 5 years of the initial PET/CT scan.Among the participants with lung cancer development, the most prevalent histologic type was adenocarcinoma (47.1%), followed by squamous carcinoma (41.2%), and small cell carcinoma (8.8%) (Supplementary Table 1).Furthermore, of the 795 participants, 585 and 210 were randomly assigned to the training and validation data sets, respectively (Fig. 2).In the training set, 100 out of 585 participants developed lung cancer within 5 years of the initial F-18 FDG PET/ CT scans and were classified into the lung cancer group.
Table 1 shows the characteristics of the participants included in the training set.In the validation set, 36 out of 210 participants developed lung cancer within 5 years of the initial F-18 FDG PET/CT scans and were classified into the lung cancer group.The characteristics of the validation sets are listed in Supplementary Table 2.

Risk prediction model
In the univariate analysis, age, sex, smoking history, lung SUVmax, SUVmean, SUVmin, SUVR, and MH index were significantly associated with lung cancer development (p < 0.001, p < 0.001, p = 0.013 p < 0.001 and p < 0.001, respectively) (Table 1).In the lung cancer group, lung SUVmax, SUVmean, SUVmin, SUVR, and MH index were significantly higher than those in the control group (p < 0.001, p < 0.001, p = 0.012 p < 0.001 and p < 0.001, respectively).The lung SUVmax, SUVmean, SUVmin, SUVR, and MH index for most of the lobes in the lung cancer group were significantly higher than those in the control group (Table 2).
All the metabolic parameters in the blood pool and liver were not significantly different between the control and lung cancer groups (Supplementary Table 5).
In the multivariate analysis, only the lung SUVR and MH index, which were better calibrated than the lung SUVmax, SUVmean and SUVmin, were considered.The clinical model included age, sex, BMI, and smoking history, and age, sex and smoking history were significantly associated with lung cancer development (age: OR 1.07, p < 0.001; male: OR 2.13, p = 0.006; past or current smoker: OR 4.60, p < 0.001).In the combined prediction model 1, all variables in the clinical model and lung SUVR were included, and age, sex, BMI, smoking history, and lung SUVR were significantly associated with lung cancer development (age: OR 1.07, p < 0.001; male: OR 2.08, p = 0.015; BMI: OR 0.93, p = 0.057; current or past smoker: OR 5.60, p < 0.001; lung SUVR: OR 1.13, p < 0.001).In the combined prediction model 2, all variables in the clinical model and lung MH index were included, and age, sex, BMI, smoking history, and lung MH index showed a significant association with lung cancer development (age: OR 1.06, p < 0.001; male: OR 1.87, p = 0.045; BMI: OR 0.93, p = 0.010; current or past smoker: OR 4.78, p < 0.001; lung MH index: OR 1.33, p < 0.001) (Table 3).

Discussion
The present study developed risk prediction models for lung cancer development using F-18 FDG PET images.The F-18 FDG uptake and metabolic heterogeneity on F-18 FDG PET images were higher in the lung cancer group than in the control group, demonstrating that chronic inflammation of the lungs is associated with the development of lung cancer.Age, sex, smoking history, BMI, and metabolic parameters were significant risk factors for lung cancer in developing prediction models.The validation result showed that these prediction models successfully distinguished between highrisk and low-risk individuals for lung cancer.The present study is the first to use F-18 FDG PET to evaluate the risk of lung cancer.
It is now evident that chronic pulmonary inflammation is involved in the tumorigenesis of lung cancer, from malignant transformation and tumor initiation to invasion and metastasis [5].Inflammatory airway injuries resulting from smoking, a variety of occupational agents, or indoor and outdoor air pollution induce cell proliferation, apoptosis resistance, inflammation, and DNA alterations that lead to lung cancer [14].Chronic pulmonary infections and inflammatory lung diseases such as asthma, chronic obstructive pulmonary disease, and fibrosis are also associated with an increased risk of lung cancer [15].However, since not all individuals with these risk factors develop lung cancer via chronic pulmonary inflammation, there is a limit to predicting lung cancer development with these risk factors based on epidemiological data.From this point of view, F-18 FDG PET can provide information on the risk of lung cancer based on the fact that the pathophysiology of lung cancer is associated with chronic pulmonary inflammation [16].This is because activated neutrophils and macrophages in the lung promote increased F-18 FDG uptake with increasing pulmonary inflammation [17].Several studies have demonstrated the usefulness of F-18 FDG PET in measuring inflammatory activity in various organs, such as the lungs, arteries, bones, and intestines [7].In the present study, it was observed that all the metabolic parameters, including the SUVmax, SUVmean, SUVmin, SUVR, and MH index, in the RLL or LLL were significantly higher compared to those in the RUL or LUL.These findings are believed to be attributed to the impact of inflammatory lung diseases, such as COPD and emphysema, which predominantly affect the lower regions of the lungs, leading to the destruction of the alveoli in those areas [18].Therefore, these findings suggest that the metabolic parameters based on F-18 FDG PET reflect the inflammatory activities in the lungs effectively.Furthermore, in the present study, we found that all metabolic parameters were higher in the high-risk individuals for lung cancer than in those with a low-risk, and demonstrated that the metabolic parameters are associated with an increased risk of lung cancer development.
Various semiquantitative parameters can be evaluated using F-18 FDG PET, including the SUVmax, SUVmean, SUVmin, SUVR, and MH index [19].The SUV, defined as  the F-18 FDG concentration at a certain time point normalized to the injected dose per unit body weight, has been generally used to measure the degree of pulmonary inflammation [8].However, despite its proven role in the diagnosis and staging of lung cancer or other malignancies, it has limitations in consistently and accurately evaluating the metabolic rate from person to person because it is a static imaging procedure measured at a single time point and is affected by various patient-related and technical factors [20].
In particular, the semiquantitative parameter using SUV to measure lung parenchyma with a low metabolic rate, rather than measuring a target with a high metabolic rate, such as a tumor, can greatly affect results even with a small error in the comparison of the metabolic rate [21].Furthermore, in the present study, the lung SUV values were often on a small scale, and their mean values yielded high odds ratios in logistic regression analyses.When an independent variable has a small scale, such as the lung SUVmax, SUVmean, and SUVmin, even a slight change in their values can lead to substantial odds ratio changes [22].However, it is important to note that such large odds ratios may not necessarily reflect realistic variable changes, as actual changes in practice tend to be small.To address these limitations associated with SUV, we considered using the lung SUVR.The SUVR, defined as the target-to-blood pool F-18 FDG uptake ratio, has been frequently used in neuroreceptor imaging and in oncology, and has been recognized as an alternative to SUV [23].The use of SUVR eliminates the shortcomings of the SUV because SUVR is a dimensionless quantity that is neither affected by scanner inaccuracies nor by erroneous body weight.In addition, the most relevant improvement in SUVR over SUV is that it can reduce the variability caused by interscan variations in the input function [24].It has been reported that SUVR exhibits a much better linear correlation with the absolute metabolic rate of glucose consumption than SUV [24].Furthermore, increasing SUVR by 1 unit does not result in excessive increases in the odds.
In the present study, lung SUVR was used to develop prediction models and showed better performance than lung SUVmax, SUVmean and SUVmin in predicting lung cancer development.In addition, prediction models using the MH index showed good performance in predicting lung cancer development.The MH index, a measure of F-18 FDG uptake heterogeneity in lung parenchyma, correlates with the extent of pulmonary inflammation and offers advantages over SUV in overcoming its limitations [25].Spatially heterogeneous F-18 FDG uptake has been associated with heterogeneous neutrophilic infiltration in the lungs [26].
Since lung cancer screening with annual low-dose CT causes additional medical costs due to the high number of false-positive screens and harmful effects of excess radiation exposure, many lung cancer prediction models using various epidemiologic data have been proposed [27].In addition, molecular and genetic markers have been incorporated to improve the performance of risk prediction models, although the observed gains have been mostly incremental.Spitz et al. [28] reported that the prediction models for smoking history performed better when adding two host DNA repair capacity markers (AUC = 0.68).El-Zein et al. [29] extended the prediction models using smoking history by adding a cytokinesis-blocked micronucleus assay endpoint, resulting in substantially improved prediction (AUC = 0.61-0.92).However, these risk prediction models using genetic biomarkers are still limited in practice, and modest gains in discrimination from incorporating genetic biomarkers would not outweigh the added costs [30].Recently, Tammemagi et al. [31] developed a lung cancer risk prediction model that included low-dose CT screening results in NLST data and demonstrated good discrimination (AUC = 0.761) in external validation data from the American College of Radiology Imaging Network (ACRIN).However, there is still a limitation, as it requires results from three chest CT scans.The present study developed novel combined prediction models using metabolic parameters measured from F-18 FDG PET and clinical factors such as age, sex, smoking history, and BMI.The combined prediction model showed higher performance (AUC = 0.867 and 0.901) than the clinical model (AUC = 0.798).Other than the clinical prediction model using age, sex, smoking history, and BMI, several clinical prediction models were developed using conventional risk factors, including race/ethnicity, education, chronic obstructive pulmonary disease, history of cancer, family history of lung cancer, and asbestos exposure [30].Thus, a prediction model combining these conventional risk factors and metabolic parameters could show better performance than the combined prediction model for lung cancer development in the present study.
There are some potential limitations in the present study.First, the study population was relatively small, based on hospital data, and was drawn only from South Korea; therefore, the results may vary in other geographic locations.In addition, the prevalence of lung cancer observed in the present study was comparatively higher than in previous studies [11].This disparity can be attributed to the exclusion criteria employed in the present study, which specifically focused on individuals who underwent follow-up chest CT or PET/CT scans.Consequently, a significant proportion of individuals who did not undergo these follow-up scans were excluded from the control group.This limitation arose from the retrospective nature of the study and that it utilized F-18 FDG PET/CT for the health screenings.To accurately assess the performance of our predictive model, further investigations are warranted, encompassing the entire study population with comprehensive follow-up.Second, the metabolic parameters on F-18 FDG PET images were measured based on manually drawn ROIs, which might cause reproducibility and harmonization problems in the present study.The larger three-dimensional ROI is largely affected by the interference of pulmonary vascular metabolism and the motion artifact of the lung, which may measure the metabolic parameters inaccurately [32].In the present study, these could be measured consistently by placing small two-dimensional ROIs on the same area in each lung lobe, far enough inward from the outer pulmonary border, less affected by chest movement.Measuring metabolic parameters by automatic segmentation of lung parenchyma, excluding pulmonary nodules and vascular structures, could be used to construct a lung cancer prediction model with better accuracy and reproducibility.Furthermore, we were unable to incorporate additional environmental and genetic risk factors into the prediction model of lung cancer development in addition to the metabolic parameters.These additional factors could enhance the reliability of the present study and improve the performance of the lung cancer prediction model.Further study is needed incorporating these additional factors, to develop the lung cancer prediction models.Third, F-18 FDG PET is not routinely performed for cancer screening.However, since the prediction model that combines metabolic parameters with clinical risk factors showed high performance in predicting lung cancer development, application of the model using F-18 FDG PET might result in a reduction in the number of individuals without lung cancer who are subjected to additional and more invasive procedures to rule out a lung cancer diagnosis following a nondiagnostic bronchoscopy [30].Even if the cancer detection rate is low on F-18 FDG PET for cancer screening [33], the benefit of F-18 FDG PET can be increased by providing additional information on lung cancer risk.

Conclusion
The metabolic parameters on F-18 FDG PET are related to an increased risk of lung cancer.The metabolic parameters on F-18 FDG PET can be used as a biomarker to provide information that is independent of clinical parameters for lung cancer risk.The prediction model that combines both clinical factors and metabolic parameters improves the predictive value for lung cancer development over the clinical model alone.Further studies with a larger number of study populations using automatic quantification of metabolic parameters on F-18 FDG PET are needed to achieve more consistent and improved results for the prediction of lung cancer development.

Fig. 1
Fig.1Representative images of regions of interest (ROIs) for calculating the metabolic parameters.Two-dimensional ROIs were drawn on the blood pool of the aortic arch and the pulmonary parenchyma in the right upper lobe and left upper lobe of the lung (a, b).The ROIs also were drawn on the pulmonary parenchyma in the right middle lobe, right lower lobe, and left lower lobe of the lung (c, d)

Fig. 2
Fig. 2 Flow diagram of the training and validation data sets.PET positron emission tomography, CT computed tomography

Fig. 4
Fig. 4 Calibration of the lung cancer prediction models tested on the validation set.a Calibration of clinical model, b Calibration of lung SUVR, c Calibration of lung MH index, d Calibration of combined

Table 2
Comparison of metabolic parameters between the control and lung cancer groups RUL right upper lobe, RML right middle lobe, RLL right lower lobe, LUL left upper lobe, LLL left lower lobe, SUVmax maximum standardized uptake value, SUVmean mean standardized uptake value, SUVmin minimum standardized uptake value, SUVR standardized uptake value ratio, MH index metabolic heterogeneity index

Table 3
Multivariable logistic regression model for the prediction of lung cancer development in the training set OR odds ratio, AUC area under the receiver-operator curve, SUVR standardized uptake value ratio, MH index metabolic heterogeneity index Clinical model Standardized β OR (95%CI) p value 0.077 (0.061-0.094)Fig. 3 ROC curves for the clinical model, simple model with SUVR, simple model with the MH index, combined model 1 with clinical factors and SUVR, and combined model 2 with clinical factors and the MH index.ROC receiver operating characteristic, SUVR standardized uptake value ratio, MH metabolic heterogeneity

Table 4
Performance of the lung cancer prediction modelsOR odds ratio, AUC area under the receiver-operator curve, SUVR standardized uptake value ratio, MH index metabolic heterogeneity index