This retrospective study was approved by our institution's internal ethics review board. Informed consent was waived for review of patient records and images. We reviewed patients who underwent surgery at a single institution between January 2018 and December 2019, and found 248 patients who underwent preoperative CT. Of these, a total of 157 patients with 159 nodules were included.

The inclusion criteria were as follows (Fig. 1): (I) histologically diagnosed adenocarcinoma, (II) CT was performed with a 1024 matrix and 0.25-mm thickness, (III) clinical stage I or II lung cancer, (IV) no previous treatment, (V) age 20 years or older.

Flowchart of patient selection

Of the 159 nodules, 115 were histologically diagnosed with adenocarcinoma. Of the 115 nodules, four with atypical adenomatous hyperplasia and four with mucinous adenocarcinoma were excluded because they differed from non-mucinous adenocarcinoma in genetic factors and cancer genesis. 21 patients (21 nodules) were excluded: patients with nodules too small to evaluate, loss of raw CT data and inability to calculate radiomic features, and those who underwent CT more than 3 months before surgery.

A total of 64 nodules with minimally invasive adenocarcinoma or invasive adenocarcinoma were finally included in our study.

All CT images were acquired on an HSR-CT scanner (Aquilion Precision: Canon Medical Systems) with 160 detector rows and 1792 detector channels. The CT parameters were as follows: tube voltage, 120 kVp; tube current, auto-exposure control; focus size, 0.6 × 0.6 mm; gantry rotation time, 0.5 s in spiral mode. Clinical images were acquired during breath-hold with full inspiration. All CT images were reconstructed with the following settings: matrix size, 1024 × 1024; slice thickness, 0.25 mm; slice interval, 0.25 mm; field of view, 34.5 cm; FC 51 (a lung algorithm) with adaptive iterative dose reduction. Volume CT dose index was 13.65 mGy ± 2.44.

All pathologic specimens were stained with hematoxylin–eosin and evaluated by pathologists at our institution according to the multidisciplinary adenocarcinoma criteria [2]. The extent of all five growth patterns (lepidic, acinar, papillary, micropapillary, and solid) was recorded by the percentage.

Commercially available software (WatchinGGO; LISIT, Co., Ltd., Tokyo, Japan) was used to segment pulmonary nodules. This software was modified for texture analysis. It can calculate radiomics features of selected two-dimensional (2D) CT images with a 1024 matrix. A total of 61 radiomic features were calculated in the segmented CT images using the software. The 61 radiomics features are listed in Online Resource 1.

Semi-automatic segmentation was performed in the maximum cross-sectional image of each tumor. The maximum cross-sectional image of each tumor was selected by a chest radiologist (MY). In the case of part-solid GGN lesions, the cross-sectional image with the maximum size of the solid component was selected. For each tumor, two independent radiologists (MT and YS) delineated regions of interest on the axial images (Fig. 2). Then, 61 radiomics features were calculated for each segmentation using our software. The average value of the radiomics features by two independent radiologists was statistically analyzed.

a A pulmonary nodule is manually surrounded by a rectangle. b The nodule is automatically segmented, and the ROI is manually corrected by the two radiologists as necessary. The radiomic features within the ROI are automatically calculated

Two image review sessions were performed. In the first session, two chest radiologists (R1 (NK) and R2 (YY)) independently classified the 64 nodules into two groups (score = 1: nodules with solid and/or micropapillary components, and score = 0: those without solid or micropapillary components). They were given the following reference information beforehand; micropapillary and solid patterns are more likely to be: (a) larger (total size≧2.5 cm), and/or (b) solid component predominant, and/or (c) spiculation or lobulation, whereas lepidic pattern is likely to contain a greater proportion of ground-glass opacities and/or an air bronchogram [7]. In the second session, performed 1 month after the first session, the same 64 nodules were reclassified into the two above-mentioned groups on the basis of the radiomics results which were significant indicators to predict the score = 1. At this time, they were also presented with the AUC value.

Statistical analyses were performed using commercially available software (MedCalc® Statistical Software version 20.216, Ostend, Belgium). Python 3.8.6 and scikit-learn 0.24.1 were used for the radiomics analysis.

The 61 radiomic features were filtered by least absolute shrinkage and selection operator (LASSO) regression with tenfold cross-validation 10 times. We determined feature importance by counting the number of times the feature had non-zero regression coefficients through repeated cross-validation. We created the radiomics features using a linear combination of the selected features. A radiomics score was calculated for each patient using a linear combination of the selected features, weighted according to their coefficients. For each higher relevant feature, the cutoff value that yielded the largest difference in the number of patients with and without solid and micropapillary components was determined using the receiver operating characteristic (ROC) method. Optimal cutoff values were determined for each variable separately using the Youden index (the maximum value of sensitivity and specificity). Associations between solid and micropapillary components and each binary group (score = 0 or score = 1) designated by the cutoff value for the ten radiomics features were evaluated by univariate logistic regression analysis. Significant features identified by univariate analysis were included in multiple logistic regression (stepwise method; P value of 0.05 or less was used for entry into the model and P value greater than 0.1 was selected for removal).

Diagnostic performance for predicting solid and micropapillary components was analyzed using ROC curves: sensitivity, specificity, and area under the curve (AUC). We used the “Comparison of ROC curves” of the MedCalc software to test the difference in the AUCs between predictive performance with and without radiomics features. Comparisons of accuracy, sensitivity, and specificity were performed using McNemar’s test. P value < 0.05 was considered significant.