Value of ¹⁸F-FDG PET/CT-based radiomics model to distinguish the growth patterns of early invasive lung adenocarcinoma manifesting as ground-glass opacity nodules

In this retrospective single-center study, we enrolled 205 patients with GGN who underwent ¹⁸F-FDG PET/CT in our department and later received surgical resection from October 2011 to October 2019. The classification of surgical pathology is based on the 2011 classification of lung adenocarcinoma published by IASLC/ATS/ERS [4]. This study was approved by the institutional ethics committee for retrospective analysis and did not require informed consent. Inclusion criteria: (1) stage I lung adenocarcinoma; (2) lung nodules manifested as GGN; (3) lesion size ≤ 4 cm; (4) PET/CT examination before surgery; (5) radical resection on tumor; and (6) PET/CT and surgery were completed within 1 month. Exclusion criteria: (1) diameter of GGN > 4 cm; (2) lesion with poor image quality or low FDG uptake that were difficult to measure; (3) patients who had received anti-tumor treatment; (4) lung adenocarcinoma stage > I; (5) history of severe liver disease, diabetes, or cancer; (6) postoperative pathological subtypes of atypical adenomatous hyperplasia, adenocarcinoma in situ, or minimally invasive adenocarcinoma; (7) unclear growth patterns or rare growth patterns (such as micropapillary and solid types); and (8) PET images did not have enough voxels (64 voxels) required by the software or metabolic volume after segmentation lower than 2.5 ml. The patient selection process was shown in Fig. 1.

Image acquisition protocols were described according to the Imaging Biomarker Standardization Initiative (IBSI) Reporting Guide [15]. All the program details were described in the electronic supplementary material 1. Within 1 month before surgery, the patients received an ¹⁸F-FDG PET/CT examination (Biograph mCT 64, Siemens, Erlangen, Germany). Based on the European Association of Nuclear Medicine (EANM) guideline 1.0 (version 2.0 was released in February 2015) [16], the ¹⁸F-FDG PET/CT images were acquired at 60 ± 5 min after ¹⁸F-FDG injection. All PET/CT images were reconstructed on a processing workstation (TureD software, Siemens Healthcare). CT data were used to perform attenuation correction on PET image, and the corrected PET image was fused with the CT image.

The LIFEx software (version 5.10, http://​www.​lifexsoft.​org) was used to extract the texture features of PET/CT images from the VOI of the lesions [17]. The patients’ PET/CT images in DICOM format were imported into the software. For PET images, experienced diagnostic physicians used the 40% and 70% threshold of SUV_max to semi-automatically set the target area of the lesion [18]. The VOI on the CT images was manually delineated and segmented slice-by-slice. The VOI covered the whole lesion, and large vessels and bronchus were excluded from the volume of the nodule. Considering the effect of different quantization levels on PET texture features, we set different higher bound of SUV (10 vs. 20) in the absolute resampling method. Finally, the software program automatically calculated and extracted 52 PET radiographic features and 49 CT radiographic features, which were provided in the supplementary material 1.

In this study, the number of radiomic features was large, but the number of cases was relatively small. To avoid model overfitting, we first used the Mann-Whitney U test to preselect the features with significant differences between acinar-papillary group and lepidic group (p value relaxed to < 0.10). Then, the least absolute shrinkage and selection operator (LASSO) algorithm was used to select the best features among the preselected features [19]. The LASSO algorithm added an L1 regularization term to the least-squares algorithm to avoid overfitting. It shrinks some coefficients and reduces others to exactly 0 via the absolute constraint. A model was generated using a linear combination of selected features that were weighted by their respective LASSO coefficients; the model was then used to calculate a radiomics signature score (rad-score) for each GGN based on the selected discriminating radiomic features. The receiver operating characteristic (ROC) curve and the area under the curve (AUC) were used to evaluate model performance.

Continuous variables were expressed as mean ± standard deviation (SD) or median (25th to 75th percentiles), and categorical variables were expressed as frequency (%). Independent t tests or Mann-Whitney U tests were used to compare continuous variables, and the Pearson chi-square test and Fisher’s exact test were used to comparing categorical variables. Multi-factor logistic regression was used to establish the prediction model, and the most optimal model parameters were selected using the minimum Akaike’s information criterion (AIC). The Bootstrap resampling method (times = 500) recommended by the TRIPOD Reporting Specification [20] was used to internally validate the model and calculate the 95% confidence interval (CI) of the AUC. A correlation heat map between each selected feature was established using the Spearman rank correlation method. The nomogram of the model was drawn in order to visualize the prediction results of each patient. A calibration curve was also drawn to show the prediction accuracy of the nomogram. ROC curve was made for each model, and the AUC of different models were compared using the DeLong method [21]. The clinical effectiveness of the model was quantified and compared using the decision curve analysis (DCA) method, which evaluates the relative cost of false positives and false negatives based on threshold probabilities. By subtracting the proportion of false positives from the proportion of true positives, and weighing the relative cost of false positives and false negatives, we can get a net benefit. The following formula was used to calculate the net benefit of model-based decisions:

Where n is the total number of patients in the study, and Pt is the given threshold probability. All analyses were performed using R3.4.3 (http://​www.​R-project.​org; software packages: glmnet, pROC, rms, dca. R). P < 0.05 was considered statistically significant. The patients with missing key parameters were excluded from the analysis, and their data were not estimated.

Finally, this study included 91 patients with IAC (23 male and 68 female), with an average age of 61.8 ± 8.6 years, ranging from 38 to 80 years. Thirteen (14.3%) patients had a history of smoking. Among the 91 patients, 59 had solitary GGN, and 32 had multifocal GGN (total lesion number 173, median lesion number 3, ranging from 2 to 36). According to the IASLC/ATS/ERS adenocarcinoma classification and prognosis standard [6, 7], 93 GGNs were classified and divided into low-risk lepidic group (n = 18), and intermediate-high risk acinar-papillary group (n = 75, 65 acinar and 10 papillary).

The acinar-papillary group had significantly higher CT_GGO than the lepidic group (P = 0.014), and the lobulated edges were also more common in the acinar-papillary group (P = 0.022). The comparison of conventional PET/CT parameters between the two groups was shown in Table 1.

The mean tumor volume segmented by the semi-automatically thresholding method (70 and 40% of SUV_max) was 6.7 ± 2.3 ml (range 2.8–12.5 ml) and 7.2 ± 2.9 ml (range 3.0–17.0 ml). On the CT images manually delineated and segmented slice-by-slice, the mean tumor volume was 6.2 ± 4.7 ml (range 0.8–20.3 ml).

Under the higher bound of SUV 20, we compared the effects of 70 and 40% delineation thresholds (PET: 64 bins from 0 to 20) on PET preselected features. It was found that compared with the 40% threshold, although 70% threshold preselected more features (22 vs. 18), the PET score produced by the 40% threshold showed better discrimination (AUC = 0.735 vs. 0.707). Besides, we also found that SHAPE_Sphericity and GLZLM_ZLNU in PET features were robust to different thresholds (Supplementary Material 2).

At 40% delineation thresholds, we compared the effects of different higher bound of SUV (10 vs. 20) on PET preselected features. It was found that the preselected features did not change significantly (especially for conventional indices, first order features, and GLCM). The PET score produced by the higher bound of 20 showed better discrimination (AUC = 0.735), but the difference from the higher bound of 10 was very small (AUC = 0.712). Besides, we also found that SHAPE_Sphericity in PET features was robust to different higher bound of SUV (Supplementary Material 2). Therefore, the final PET parameters were 40% thresholds and 64 bins from 0 to 20, while CT used the default parameters.

LASSO algorithm and 10-folds cross-validation were used to extract the best subset of radiomic features, and four radiomic features were extracted (Fig. 2), which were SHAPE_Sphericity, GLZLM_ZLNU, HISTO_Kurtosis, and GLZLM_SZLGE. Among the above features, the first 2 are PET features, and the last 2 are CT features. The following formula was used to calculate the rad-score for each GGN:

Table 2 shows the median and interquartile range of the four selected radiomic features and the calculated rad-score. There were significant differences in rad-score and the four selected features between the lepidic group and the acinar-papillary group (all P < 0.05).

Figure 3 shows the correlation heat map of preselected PET/CT features, which illustrates the results of feature selection. The color keys and histogram bars in the upper left corner indicate the correlation between each image feature. A correlation equals to 0 indicates the best independence among the corresponding features, while a correlation equal to 1 or − 1 suggests a perfect correlation. We found that the four extracted PET/CT features were independent of each other (|r| all < 0.5), indicating that these features could convincingly represent the tumor features and the prediction model was reliable.

We combined rad-score and conventional CT parameters (edge and CT_GGO) to establish a multivariate logistic regression model (joint model) and used a non-parametric resampling method (Bootstrap resampling, times = 500) to perform internal verification.

The model is as follows:

The nomogram and a calibration curve of the joint model were drawn (Fig. 4a, b). There was good consistency between the predicted and observed values, and the ROC curve of the joint model showed an AUC of 0.804 (95% CI [0.699 – 0.895]) (Fig. 4c).

To evaluate the performance of radiomic features in predicting GGN growth patterns, we compared the rad-score, CT_GGO, and the joint model using ROC (Fig. 5). The prediction capabilities of the three models are listed in Table 3, including AUC, sensitivity, specificity, accuracy, positive likelihood ratio, and negative likelihood ratio. The results showed that the AUC of the joint model and rad-score were higher than CT_GGO (0.804 vs. 0.675 and 0.790 vs. 0.675), but the difference was not statistically significant (P = 0.109 and 0.132). There was also no significant difference between the joint model and the rad-score (P = 0.605).

Since the AUC of the joint model, rad-score, and CT_GGO were not significantly different, we introduced DCA in order to evaluate the performance of the three models (Fig. 6). Under the purpose of screening for intermediate-high risk growth patterns (sensitivity ≥ 0.800, threshold probability ranging from 0.73 to 0.98), the net benefit of the joint model was better than rad-score and CT_GGO; similarly, under the purpose of confirming the diagnosis of intermediate-high risk growth patterns (specificity ≥ 0.833, threshold probability ranging from 0.30 to 0.59), there was no significant difference in net benefit between the three. Thus, the overall clinical value of the joint model was higher than the other two.