The ethics committee of the First Affiliated Hospital of Chongqing Medical University approved this retrospective study. All methods were carried out in accordance with relevant guidelines and regulations. Informed consent (written consent) was obtained from all individual participants included in the study. In this research, F-PLC was defined as a special type of lung cancer presenting as focal consolidation involving less than half of the lobe. Meanwhile, F-PIL refers to focal inflammatory consolidation affecting less than half of the lobe. The inclusion criteria were as follows: (1) patients with a pathological diagnosis confirmed via surgical resection; (2) patients who underwent chest CT scan at our institutes; (3) CT results indicated a solitary and focal consolidation, characterized by an increased density of lung parenchyma with obscuration of the underlying vessels, with a polygonal shape (e.g., triangular, rectangular, or trapezoidal) and the largest slice involving less than half the area of a lobe on axial CT images; (4) patients with interval of < 1 month between CT imaging and subsequent pathological analysis. Meanwhile, the exclusion criteria were as follows: (1) patients who did not undergo chest contrast-enhanced CT scan; (2) patients with unsatisfactory imaging quality due to respiratory motion artifact; (3) patients who received any anti-tumor or anti-inflammatory therapy before initial chest CT scanning; (4) patients with incomplete clinical data. In total, 346 (193 men and 153 women; mean age: 61.5 ± 14 [range: 22–86] years) patients diagnosed between January 2015 and May 2021 at center 1 were included. Among them, 209 presented with F-PLC and 137 with F-PIL. We randomized the patients into two groups, with a ratio of 7:3 (242 patients in the training cohort and 104 patients in the internal validation cohort) (Fig. 1). Moreover, we included 50 consecutive patients (29 men and 21 women; mean age: 64 ± 13.5 [range: 38–87] years) diagnosed at center 2 in the external validation cohort. Among them, 27 presented with F-PLC and 23 with F-PIL.
CT image acquisition and morphological features analysis
All chest contrast-enhanced CT examinations were performed using Discovery CT750HD (GE Healthcare) and Optima CT660 (GE Healthcare). All patients underwent CT scan in a supine position at the end of inspiration during a single breath hold to prevent respiratory motion artifacts. The CT scanning parameters were as follows: tube voltage, 100–130 kVp; tube current, 100–250 mA; slice thickness, 5.0 mm; and slice interval, 5.0 mm. Nonionic iodinated contrast medium (iohexol 300 mg iodine/mL; Omnipaque, GE Healthcare) at a dose of 1.5 mL/kg was administered at a flow rate of 3.0 mL/s using a dual-high-pressure injector via the antecubital vein. Then, 50 mL of saline solution was injected. The arterial and delayed phases were triggered at 30 and 120 s, respectively.
Two thoracic radiologists (with 13 and 6 years of experience in chest imaging, respectively) reviewed all CT images in a PACS workstation (Vue PACS, Carestream) together. In cases of disagreement, a consensus was reached via discussion. The lung window (window width, 1600 HU; window level, − 600 HU) and mediastinal window (window width, 450 HU; window level, 50 HU) of CT images were used for assessment. CT morphological features including lesion size (maximum diameter of the lesion on axial CT images), margin (well-defined, with a clear border definition; ill-defined, with a partial or completely blurred border), spiculation (linear strands extending from the nodule or mass margin into the lung parenchyma without reaching the pleural surface), air bronchogram (tubelike or branched air structure within the lesion), pleural attachment (lesion attaching to the pleura including the fissure and lesion margin obscured by the pleura), necrosis (unenhanced areas with clear boundaries within the lesion), calcification, lymphadenopathy (hilar or mediastinal lymph nodes with short-axis diameter larger than 1 cm), and pleural effusion were evaluated.
Comparison of clinical and CT features between both groups
The clinical and CT morphological features including age, gender, smoking history, respiratory symptoms, lesion size, margin, spiculation, air bronchogram, necrosis, calcification, lymphadenopathy, pleural attachment, and pleural effusion between the F-PLC and F-PIL groups in center 1 were compared.
Model establishment and performance evaluation
Establishment of clinical model
The clinical model was constructed by incorporating clinical and CT morphological features that differed significantly between the F-PLC and F-PIL groups in center 1 using the multivariate logistic regression analysis.
Establishment of radiomics model
The region of interest (ROI) on non-contrast-enhanced CT images with lung window was manually segmented by one radiologist (with 6 years of work experience in thoracic imaging) who did not know the clinical and pathological information of patients using the ITK software (version 2.2.0, http://www.itksnap.org/pmwiki/pmwiki.php). Three ROIs were cautiously drawn along the margin of the lesion in the largest layer and the adjacent upper and lower layers from axial CT images, covering the whole contour of the solid component of the lesion. Figure 2 shows the radiomics analysis flowchart. To ensure consistency, these delineations were conducted three times. To evaluate intra observer repeatability, the observer repeated the ROI delineation 4 weeks after the first assessment. Thereafter, the intra-class correlation coefficients (ICCs) were calculated to evaluate the stability and reproducibility of feature extraction. These features with ICC values > 0.75 were included in this study.
Radiomics feature extraction was performed using Pyradiomic implemented in Python (https://pyradiomic.readthedocs.io/en/latest/), which can extract radiomics features from CT images with a large panel of engineered hard-coded feature algorithm. As shown in Fig. 2, the steps for selecting radiomics features and signature building were as follows: first, we imported the CT images into the software with Digital Imaging and Communications in Medicine files (planning CT images with GTV ring). The images were preprocessed and segmented, and the resolution feature matrix was normalized. Second, we compared the similarity of each feature pair. If the Pearson correlation coefficient of a feature pair was > 0.90, one of them will be deleted. Third, the optimal subset method with the Akaike information criterion (AIC) information criteria were combined to select features. Finally, we established the logistic regression radiomics model using a combination of features under the minimum AIC correspondence.
Establishment of combined model
The combined model, which integrated CT-based radiomics signatures, clinical factors, and CT morphological features, for differentiating F-PLC from F-PIL was developed via a multivariate logistic regression analysis.
Performance evaluation of the three models
The receiver operating characteristic curve (ROC) and the area under the curve (AUC), accuracy, sensitivity, and specificity were used to evaluate the performance of three models in the training and internal and external validation cohorts, respectively.
Statistical analysis was performed using the Statistical Package for the Social Sciences software for Windows (version 19.0, IBM, Armonk, NY, USA) and R software (version 3.6.1; http://www.Rproject.org). Quantitative data with a normal distribution were expressed as mean ± standard deviation, whereas those with a non-normal distribution were presented as medians ± interquartile ranges. Meanwhile, categorical variables were expressed as numbers and percentages. The Chi-square test, Two-independent-samples Student’s t-test, and Mann–Whitney U test were used in the univariate analysis. DeLong test was performed to compare the AUCs of three models in the internal and external validation cohorts. A two-sided P-value of < 0.05 was considered statistically significant.