This retrospective study was conducted after permission from the ethics committee in Shanghai Pulmonary Hospital, and the informed consent requirement of patients was waived. The inclusion criteria were as follows: (a) pathological examination confirmed lung adenocarcinoma or pulmonary cryptococcus; (b) interval between surgery and preoperative chest CT less than 2 weeks; (c) solitary pulmonary solid nodule ≤30 mm in diameter; and (d) the thickness of the latest CT images before the surgery was not more than 1.5 mm. The cases would be excluded if they meet one of the following conditions: (a) inconclusive pathological diagnosis from an inadequate biopsy tissue specimen or bronchoscopy; (b) impure-solid nodules (pure ground-glass nodules or solid nodules with ground-glass opacity); (c) the number of pulmonary nodules was greater than one or a primary nodule with a couple of scattered lesions; (d) obvious calcifications inside nodule; and (e) patients affecting malignant tumors.

Clinical baseline characteristics were collected from the hospital information system included age, gender, smoking history, and immune status. Patients that meet one of the following criteria will be considered as immunodeficient: (a) a history of immunosuppressive therapy (treatment with immunity inhibitor), (b) diabetes mellitus, (c) AIDS, (d) recipients with a transplanted organ, (f) severe respiratory system limitation, or (g) other systemic diseases (such as lupus) and receiving steroids therapy (13–15).

The chest images of the current study were acquired using one of the two CT scanners: (a) the Brilliance 40 scanner (Philips, Netherlands) with a protocol of 120 kVp tube energy and 200 mAs tube current; the parameters were 40 × 0.625 mm detector, 512 × 512 matrix, and 0.4 pitch; (b) the Somatom Definition AS scanner (Siemen, Germany) with a protocol of 120 kVp tube voltage and 130 mAs effective dose. The machine parameters were 64 × 0.625 mm detector, 1.0 pitch, and 512 × 512 matrix. The 1.0-mm thickness and 0.7-mm increment were used as reconstruction standards and applied to all CT images.

The review of the CT images of all cases was firstly completed by a junior radiologist (YY, with 6 years experience in thoracic radiology) and then inspected by a senior radiologist (XS, with more than 30 years in thoracic radiology); both were unaware of the actual pathological conclusion and clinical characteristics and analyzed the CT findings of each SPSN without discussion. All CT images were read with lung windows (level, −450HU; width, 1,500HU) and mediastinal windows (level, 40HU; width, 400HU). The following characteristics of each nodule in chest CT were assessed and recorded: (a) maximum diameter (the longest diameter of the nodule in the largest axial section); (b) size (the average of the long and short axis of the lesion on the slice, in which the nodule was the largest); (c) location (upper or middle, lower); (d) shape (round or ellipse, irregular); (e) lobulation (the contour of the nodules exhibiting concave and convex); (f) spiculation (the strands around the nodule margin and without contacting the pleural surface); (g) air bronchogram (tube-like low attenuation areas within the lesion); (h) pleural retraction (retraction of the pleura toward the nodule); and (i) cavity (lower-density shadow or air chamber in the nodule). The consensus was reached by consultation when inconsistencies between the observers exist.

Combining clinical features and subjective CT findings, the clinical model was established in two steps. Firstly, the comparison in category variables (semantic characteristics, smoking history, and immune status) between two sets used the chi-squared test or Fisher’s exact test; numerical variables (maximum diameter and size) used t-test or Wilcoxon test to assess the differences properly. Secondly, we picked up the variables whose difference between the groups was statistically significant according to the information of Akaike and established the clinical model through the stepwise backward selection of variables by multivariate logistic regression.

The nodules’ boundaries on each cross-section were traced based on an open software 3d-slicer (version 4.9.0, http://www.slicer.org) on the lung window background (level, −450HU; width, 1,500HU) by a radiologist (JZ) with 1 year of experience in chest radiology and reviewed by an expert radiologist (XS) with more than 30 years of experience in thoracic imaging. All radiologists were blind to the patient information, carefully avoiding vessels, bronchus, and pulmonary parenchyma and keeping an approximate distance of 1–2 mm from the nodule margin.

Radiomic features of all nodules were automatically extracted by an in-house software based on Python (version 3.7.0, http://www.python.org) using the pyradiomics package (16). To minimize the potential impact caused by the variability of imaging parameters and scan conditions of different CT scanners and lift the convergence speed of the classification model, all CT images were preprocessed with a series of methods, including images resampled, histogram discretion, and normalization based on Z-score before radiomics feature extraction.

To develop a radiomics signature with robustness, high relevance, and efficient performance, a three-step feature selection process was implemented in our study: (a) assessing the reproducibility and difference of radiomics features; (b) selecting the most relevant and filtering redundant radiomics features between PC groups and LAC groups; and (c) building radiomics signature with the optimal subset of features.

Firstly, to select features insensitive to segmentation for improving the reproductivity of the model, the intra-class correlation coefficients (ICCs) were chosen as the criterion for evaluation. A radiologist (KS) aimlessly picked 30 CT images (15 PC and 15 LAC) from the training cohort and manually drew the contours of nodules independently. ICC values of more than 0.75, which means good reliability, were retained for the following analysis. Secondly, in the present study, we used the max-relevance and min-redundancy (mRMR) algorithm to rank the radiomics feature according to their correlation and redundancy between PC and LAC. Ultimately, the top-ranking features were reserved for constructing radiomics signature by calculating mutual information. Thirdly, to select the combination of the most helpful features in distinguishing between PC and LAC to optimize the predictive performance of the model, the least absolute shrink and selection operator (LASSO), an approach that calculates the regression coefficients and successively shrinks them to avoid overinflation, was implemented with fivefold cross-validation for the optimization process. Furthermore, we applied the Mann–Whitney U test to examine the statistical significance between the two cohorts in radiomics signature.

To screen out the independent predictive factors for differentiating LAC from PC lesions in clinical features, radiologic signs, and radiomics signature, the multivariate logistic regression analysis was performed in the above variables. Finally, the variables with a significant statistical difference were reserved and used to construct the radiomics nomogram model.

The Hosmer–Lemeshow test was implemented to assess whether the information from variables in the nomogram was extracted and integrated by the model adequately. Furthermore, calibration curve analysis was plotted to visualize the predictive model performance. In order to evaluate the individual predictive ability of the three models, the receiver operating characteristic (ROC) curve analysis was drawn. The evaluation index of predictive performance, such as AUC, sensitivity, specificity, accuracy, NPV, and PPV, were calculated in the training and test sets, respectively. We utilized the Delong test to identify whether the difference of the AUCs from various models reaches statistical significance. Decision curve analysis (DCA) was delineated to directly present the net benefits in clinical practice with different risk thresholds. Furthermore, the net reclassification index (NRI) of the three models was calculated to compare their prediction performance.

ALL statistical analysis was performed using R statistical software (version 3.3.1), Python (version 3.7.0), or SPSS 20 (IBM, Armonk, NY, USA). The ROC analyses and LASSO were performed using the “sklearn” package in Python. Package “pymrmr” was used to execute the mRMR algorithm. The nomogram and calibration curve analysis was completed by “rms”, and DCA was completed by “rmda” in R programming. NRI and Delong test were calculated by the package “nricens” and “pROC”, respectively. The AUC cutoff values were established by calculating the maximal Youden index (sensitivity + specificity − 1). All statistical analyses were two-sided, and significance was set at p < 0.05.

Between January 2011 and December 2020 in our hospital, we enrolled 1,482 patients (213 PC and 1,269 LAC) who fulfilled the inclusion criteria. Because the sample size of PC is much smaller than that of LAC, we adopted the 1:1 propensity score matching (PSM) approach to balance the size of the number of two cases based on age and gender. Finally, 426 patients (213 PC and 213 LAC) were matched and split into the training (151 PC and 147 LAC) and test (62 PC and 66 LAC) sets in a 7:3 ratio for subsequent research. The clinicopathological factors and the recorded CT features are displayed in Table 1 .

After matching, a total of 426 patients were enrolled in the present study. The training cohort contained 298 patients—171 males (age range, 24–81 years; mean age, 56.46 ± 10.93 years) and 127 females (age range, 28–82 years; mean age, 57.11 ± 10.02 years). The test cohort had 128 patients—77 males (age range, 32–78 years; mean age, 57.98 ± 11.19 years) and 51 females (age range, 38–77 years; mean age, 53.97 ± 8.77 years). The maximum diameter, size, lobulation, and pleural retraction had significant statistical differences between the PC and LAC in the training cohort (p < 0.05, Table 1 ).

Multivariable analysis is presented in Table 2 and showed that the maximum diameter (odds ratio [OR], 1.075; 95% confidence interval [CI], 1.032–1.120; <0.001), size (OR, 1.069; 95% CI, 1.022–1.118; =0.003), lobulation (OR, 0.174; 95% CI, 0.088–0.344; <0.001), and pleural retraction (OR, 0.173; 95% CI, 0.097–0.306; <0.001) were independent predictive factors for the clinical model.

Through the inter- and intra-observer agreement test, 631 radiomics features had ICC values greater than 0.75 totally, which means good reproductivity. Then, the mRMR analysis was performed with 631 radiomics features (ICC > 0.75), and the top 100 highly relevant and nonredundant features were picked up for radiomics signature construction. The most optimal feature combination for distinguishing PC from LAC was calculated by the LASSO algorithm. When lambda (λ) was 0.0118 and log(λ) was −1.9293, a subset consisted of 24 radiomics features was screened out to develop a radiomics signature ( Figures 1A, B ). Figure 2 shows the coefficients of the 24 potential predictors in order of importance.

The result of the Wilcoxon Rank Sum test suggested that the radiomics signature had a significant statistical difference between pulmonary cryptococcosis and lung adenocarcinoma in training and test sets (all p < 0.0001), which indicated that the radiomics signature had a good identification efficiency in the training (AUC = 0.8519; 95% CI, 0.7881–0.9157) and test (AUC = 0.8306; 95% CI, 0.7853–0.8759) cohorts, respectively ( Figure 3 ).

In accordance with the multivariable logistic regression analysis (shown in Table 3 ), the radiomics signature (OR, 1,365.794; 95% CI, 174.8112013–10,670.895; p < 0.001), maximum diameter (OR, 0.904; 95% CI, 0.842–0.970; p = 0.005), lobulation (OR, 0.292; 95% CI, 0.128–0.670; p = 0.004), and pleural retraction (OR, 0.279; 95% CI, 0.142–0.549; p < 0.001) were statistically significantly different and independent differentiators. On the basis of these four independent factors, we developed a combined radiomics nomogram incorporating radiomics signature and clinical characteristics ( Figure 4A ).

The calibration curve analysis was plotted and revealed an apparent connection between the actual and predicted labels ( Figures 4B, C ). The Hosmer–Lemeshow test showed that the result was statistically nonsignificant, which means fine goodness of fit.

Table 4 presents the various indexes of diagnostic efficiency of three models in training and test cohorts, respectively. To intuitively demonstrate and compare the identification performance of these models, the ROC curves are designed in Figure 3 . When the model was generated with clinical characteristics alone, the AUC = 0.7901 (95% CI, 0.7398–0.8403) in the training set, which was inferior to another single model including the radiomics features alone in which the AUC = 0.8519 (95% CI, 0.7881–0.9157). However, when clinical characteristics and radiomics signature were integrated into a radiomics nomogram, the combined model yields the best discrimination effectiveness in the training cohort and the AUC was increased to 0.9101 (95% CI, 0.8787–0.9416).

In the test cohort, the AUC of nomogram fell slightly compared to the training cohort, but still kept a decent classification ability (AUC = 0.8881; 95% CI, 0.8336–0.9425; sensitivity = 0.8013; specificity = 0.8333; accuracy = 0.8203) ( Table 4 ). Significant differences (nomogram model vs. clinical model) concerning AUCs were confirmed by the Delong test in the training set (p < 0.0001) and test set (p < 0.0001), respectively. Likewise, the improvement of differentiation capacity of radiomics nomogram in comparison to the subjective CT model had been demonstrated by NRI (training cohort: NRI = 0.2672, p = 0.0068 and test cohort: NRI = 0.2815, p = 0.0462). Finally, the DCA indicated that the nomogram model brought more clinical net benefit to patients with a SPSN compared to the single model containing clinical features alone within most of the range of the threshold probability ( Figure 4D ).