This study enrolled small SPN patients consecutively from January 2019 to November 2021. The inclusion criteria included patients who indicated: (i) small SPNs ≤ 20 mm, (ii) a confirmed pathological SPN diagnosis after surgical resection, and (iii) a < 2-week interval between SPN detection and surgical resection. Patients were excluded if they had: (i) poor image quality; (ii) a history of malignancy, (iii) SPNs < 6 mm in diameter, or (iv) incomplete clinical data. Eligible patients were randomly assigned to training and testing cohorts at an 8:2 ratio.

Clinical data were collected for all the patients including demographic factors (age, gender, smoking history), CT features (location, diameter, lobulation, spiculation, pleural retraction sign, CT bronchus sign, and calcification), and the levels of tumor markers [carcinoembryonic antigen (CEA), squamous cell carcinoma antigen (SCC), neuron-specific enolase (NSE), serum gastrin, cytokeratin 19 fragment (CYFRA21-1)]. The size of the SPNs was measured as the largest diameter on the axial CT images.

A 64-row CT instrument (Brilliance 64 CT, Philips) was used for all CT imaging with the following settings: tube voltage = 120 kVp, tube current = 160-220 mAs, pitch = 0.97, and collimation = 0.6×128 mm. Images were reconstructed using a medium sharp (B50) reconstruction algorithm with a 1.0 - 1.25 mm thickness. The images of the lung (width = 1600 HU; level = -600 HU) and mediastinal (width = 450 HU; level = -50 HU) windows were analyzed. CT imaging features were assessed individually by two chest radiologists (JXS and XXZ) with 7 and 12 years of relevant experience, respectively, who were blinded to the pathological results for each patient.

A chief radiologist (YJY) with 7 years of experience manually segmented target 3D SPNs with the Radcloud platform (http://radcloud.cn) and remained blinded to patient pathological results. Further, the Radcloud platform was used for extracting the radiomics feature. Observer consistency was assessed using intra- and inter-class coefficient (ICC) values. Briefly, CT images from 20 randomly selected individuals in the training cohort were independently segmented by two radiologists (JXS and XXZ). Moreover, Reader 1 (JXS) repeated the segmentation of tumors from these 20 patients following a one-week interval. Repeatable features were regarded as those with an ICC ≥ 0.8, which were elected for subsequent evaluation. All remaining images were segmented by Reader 1 (JXS).

Features with > 0.8 variances were identified with the variance threshold method for further analysis. Furthermore, based on the Selec-K-Best method analysis, features with a p-value of < 0.05 were then retained for a final step in which, features associated with malignant SPNs were selected using a least absolute shrinkage and selection operator (LASSO) regression model. These characteristics were employed to formulate a radiomics signature, enabling the computation of radiomics scores for each respective patient.

To distinguish between malignant and benign small SPNs, a CT radiomics-based model was established. Briefly, univariate analysis (UA) and multivariate logistic regression analysis (MLRA) were carried out to select risk factors related to SPN malignancy in the training cohort. Then, a nomogram incorporating these risk factors and radiomics scores was established. Subsequently, the AUC values for receiver operating characteristic (ROC) curves were employed to assess the accuracy of the developed model. Moreover, the model was validated using data from the patients in the testing cohort.

The clinical application of the predictive model was assessed via a decision curve analysis of the training and testing cohorts.

The statistical analysis was carried out using the SPSS 25.0 and R 4.1.2 software. Eligible patients were randomly assigned to training and testing cohorts at an 8:2 ratio using the Radcloud platform. The comparison of categorical data was carried out via Fisher’s exact test or χ² test, while for continuous data, an independent sample t-test or Mann-Whitney U test was carried out. UA and MLRA were performed to identify factors that are associated with SPN malignancy. In the MLRA, particularly in UA, variables indicating a P-value < 0.1, were then selected. The comparison of AUC values was implemented using the DeLong test, and the predictive model’s performance was evaluated via calibration curves and the Hosmer-Lemeshow test. The statistical significance threshold was set as P < 0.05.

This study recruited 323 patients with small SPNs who underwent surgical resection procedures in our hospital from January 2019 to November 2021. Of these 323 patients, 210 were selected for further analyses ( Figure 1 ). Each patient had a single SPN with a pathological diagnosis confirmed following surgical resection. All the participants were categorized into training (n = 168) and testing (n = 42) groups in an 8:2 ratio. Table 1 indicates detailed information on the characteristics of selected patients.

Initial analyses identified 1409 radiomics features. Then, to develop a radiomics score, a stepwise process was then employed ( Supplementary Table S1 ), which revealed 9 features for radiomics score calculation ( Supplementary Table S2 ). Coefficient values for all features as well as the mean square error for the combined sequences are presented in Supplementary Figure S1 .

The clinicoradiological features associated with malignant SPNs were assessed in the training cohort. The data revealed that the training cohort comprised 88 and 80 malignant and benign SPNs patients, respectively. UA identified older age (P = 0.01), lobulation (P < 0.001), spiculation (P < 0.001), and larger SPN diameter (P < 0.001) as being associated with a risk of SPN malignancy. Furthermore, MLRA confirmed that lobulation (P < 0.001), spiculation (P < 0.001), and larger SPN diameter (P < 0.001) were associated with a greater risk of SPN malignancy ( Table 2 ).

The identified clinicoradiological factors and radiomics score were used to establish a predictive model with the following formula: X = -6.773 + 12.0705×radiomics score+2.5313×lobulation (present: 1; no present: 0)+3.1761×spiculation (present: 1; no present: 0)+0.3253×diameter. A nomogram was also developed with this CT radiomics-based model ( Figure 2 ). Furthermore, individual clinicoradiological and radiomics score models were also developed. Sensitivity, specificity, accuracy, and AUC measurements for these models are presented in Table 3 . AUC values for the CT radiomics-based model, CT radiomics score, and clinicoradiological score were 0.957, 0.945, and 0.853 ( Figure 3A ). Moreover, the AUC values of the CT radiomics-based model were significantly higher those for both CT radiomics scores (P = 0.035) and clinicoradiological scores (P = 0.021).

The testing cohort comprised 42 patients, including 20 benign and 22 malignant SPN patients. Using the models developed above as well as the testing cohort data, the AUC values of the CT radiomics-based model, CT radiomics score, and clinicoradiological score were assessed as 0.943, 0.916, and 0.816, respectively ( Figure 3B ). The AUC of the CT radiomics-based model was significantly greater than the CT radiomics score (P = 0.043) and clinicoradiological score (P < 0.001).

In calibration curve analyses, the results predicted via the CT radiomics-based model and the actual results indicated good consistency in the training and testing cohorts ( Figures 4A, B ). Further, decision curves confirmed that the developed nomogram and associated predictive model yielded a net benefit in both cohorts with a risk threshold > 0 ( Figures 5A, B ).