Diagnostic performance of spectral CT quantitative parameters in predicting spread through air spaces of lung adenocarcinoma Authors: Weicong Chen, Ziyang Xia, Jinyang Ma, Wang Liu, Bohang Cui, Shuchan Zheng, Qi Wan, Qianyi Li, Yuting Wu, Tao Chen, Tao Ye, Yeqing Zhu, Xinchun Li Published: 2026 ABSTRACT BackgroundSpread through air spaces (STAS) is associated with adverse outcomes in lung adenocarcinoma (LUAD). Although conventional computed tomography (CT) features have shown predictive value for STAS, the incremental diagnostic value of spectral CT quantitative parameters remains unclear. This study aimed to evaluate the diagnostic performance of spectral CT in predicting STAS in solid and part-solid nodules.MethodsA total of 164 patients with surgically confirmed LUAD presenting as solid and part-solid nodules who underwent dual-phase spectral CT scans were included. Conventional CT features (e.g., consistency, margin, spiculation) and spectral CT parameters [e.g., monochromatic CT values at 100 keV, normalized iodine concentration (NIC), effective atomic number (Zeff), and energy attenuation curve slope (λ)] were extracted. Three multivariate logistic regression models: conventional CT model, spectral CT model, and an integrated model were constructed. Model performance was assessed using the area under the curve (AUC), and AUCs were compared using DeLong test.ResultsIn the multivariable study, lobulation [odds ratio (OR) =3.509, 95% confidence interval (CI): 1.059–11.627], whole tumor size (OR =1.060, 95% CI: 1.013–1.110), and consolidation-to-tumor ratio (CTR; OR =72.057, 95% CI: 14.619–355.176) were significant conventional CT predictors, whereas arterial λ (OR =0.578, 95% CI: 0.396–0.846) and CT number at venous 100 keV (OR =1.017, 95% CI: 1.009–1.024) were significant spectral CT predictors of STAS positivity. The integrated model achieved the numerically highest AUC (0.842, 95% CI: 0.777–0.894), followed by the conventional CT model (AUC =0.829, 95% CI: 0.763–0.883) and the spectral CT model (AUC =0.829, 95% CI: 0.762–0.883). However, the differences among the three radiological models were not statistically significant (all P>0.05).ConclusionsSpectral CT parameters provided diagnostic performance comparable to that of conventional CT features for predicting STAS in solid and part-solid nodules. However, adding spectral CT parameters to conventional CT features did not yield a statistically significant incremental diagnostic value in this cohort. Introduction Lung cancer remains the leading cause of cancer-related mortality worldwide, with lung adenocarcinoma (LUAD) being the most prevalent histological subtype (1). The World Health Organization (WHO) first officially established spread through air spaces (STAS) in 2015 (2), which was defined as the spread of tumor cells in the airspace beyond the margins of the primary tumor. It has since been recognized as an independent adverse prognostic factor associated with higher risk of recurrence, lymph node metastasis, and poorer survival rate (3-5). A recent meta-analysis found no significant difference in overall survival (OS) between sublobar resection and lobectomy for patients with early-stage lung cancer (6). However, in STAS-positive patients, lobectomy was associated with better outcomes than sublobar resection (5). Therefore, preoperative assessment of STAS is critical for surgical decision-making. Intraoperative frozen section is a highly specific but not sensitive method for STAS detection, with a further limitation of moderate inter-reader and intra-reader agreement overall (7). Previous studies have indicated that several conventional computed tomography (CT) morphological features are significantly associated with STAS (8-10). A recent meta-analysis demonstrated that conventional CT models have fair-to-considerable diagnostic efficacy for STAS prediction [pooled area under the curve (AUC): 0.80, 95% confidence interval (CI): 0.76–0.83] (11). Spectral CT acquires X-ray attenuation data across different energy spectra, enabling material differentiation and quantitative analysis of tissue composition (12,13). It has shown clinical value in lung cancer, including differentiation between benign and malignant lesions, assessment of the biological behavior of pulmonary nodules, and evaluation of lymph node metastasis (14-16). A previous spectral CT study suggested added value for STAS prediction (17), but it excluded several histologic subtypes according to postoperative pathology and included pure ground-glass nodules (GGNs). These selection criteria are less consistent with preoperative decision-making, because postoperative histology is unavailable before surgery and STAS is uncommon in pure GGNs (10,17,18). This study aimed to evaluate the potential of three radiological models using spectral CT and/or conventional CT features in predicting STAS preoperatively in solid and part-solid pulmonary nodules. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0705/rc). Methods Study participants The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of The First Affiliated Hospital of Guangzhou Medical University. The requirement for informed consent was waived based on the retrospective nature of the study. All consecutive patients who were suspected of lung cancer and underwent surgery at our institution from 1 January 2025 to 10 November 2025 were reviewed. The inclusion criteria were as follows: (I) patients who underwent surgical resection with postoperative histopathological confirmation of LUAD; (II) patients who underwent dual-phase spectral CT scans within two weeks prior to surgery; and (III) patients who had part-solid or solid nodules on preoperative CT. Patients were excluded according to the following criteria: (I) pure GGNs on preoperative CT; and (II) poor CT image quality due to respiratory motion artifacts. The patient selection process is illustrated in Figure 1. Flowchart of patients’ selection. CT, computed tomography; pGGN, pure ground-glass nodule; STAS, spread through air spaces. Clinical and pathological characteristics Clinical information included patient demographics (sex and age), smoking status, surgical approach, preoperative serum tumor marker levels [neuron-specific enolase (NSE), cytokeratin fragment 21-1 (CYFRA21-1), carcinoembryonic antigen (CEA), and squamous cell carcinoma antigen (SCCA)], and tumor, node, metastasis (TNM) stage was obtained from the electronic medical records and chart review. Pathological diagnosis was extracted from postoperative histopathological reports which were reviewed and finalized by specialized pathologists on chest with over 10 years of experience. CT scan parameters All dual-phase spectral CT scans were performed using a 256-row CT scanner (Revolution Apex, GE Healthcare, Chicago, IL, USA). Scanning parameters were as follows: high and low energy tube voltage (80 and 140 kVp) with fast switching, automatically modulated tube current, collimation width of 0.5 mm, field of view (FOV) of 400 mm × 400 mm, matrix size of 512×512, a pitch of 0.992, and a rotation time of 0.5 s. The scanning range included the entire chest from the lung apices to the bases in a single breath-hold at maximum inspiration. Unenhanced scanning using the Gemstone Spectral Imaging mode was performed first. Patients were then injected with iodinated contrast agent (Ultravist 300 mg/mL; Bayer AG, Leverkusen, Germany) at 1.2–1.3 mL/kg of body weight by using a high-pressure injector (Ulrich GmbH & Co.KG, Ulm, Germany) at a flow rate of 3.0 mL/s, followed by a 40 mL saline flush at the same flow rate. Arterial and venous phase scans were carried out at 20 and 50 s after the start of contrast agent injection, respectively. After scanning, conventional CT images and spectral-based images were reconstructed at 0.625-mm slice thickness at 0.625-mm intervals. Conventional CT imaging features On conventional CT images [window width, 1,500 Hounsfield unit (HU); window level, −650 HU], morphological features including nodule location, consistency (part-solid or solid), margin (smooth or non-smooth), shape [regular (round or oval) or irregular], spiculation, lobulation, pleural retraction sign, bubble-like lucency sign (spots of air attenuation or low density similar to normal lung tissue within a nodule), air bronchogram sign (dilated or disrupted small bronchiole passing through the nodule), nodule size (average of the maximum length and maximum width defined as the perpendicular to the length on the same axial image showing the maximum area of the entire nodule), solid component size (average of the maximum length and maximum width defined as the perpendicular to the length on the same axial image showing the maximum area of the largest solid component), and consolidation-to-tumor ratio (CTR) were evaluated and recorded. Each chest CT image was reviewed by two chest radiologists (W.C. and B.C., with 3 and 7 years of experience, respectively). Upon completion, any disagreement on the characteristics of each nodule was jointly reviewed, and in case of disagreement, the final classification was made by consensus of two senior radiologists (Y.Z. and X.L., both with 10+ years of experience in chest radiology). All radiologists were blinded to patients’ clinical data and histological diagnosis. Spectral CT analysis Monochromatic images at 40 and 100 keV both in the arterial phase and venous phase, iodine-based material decomposition maps, and effective atomic number (Zeff) maps were generated by Advanced Workstation 4.7 (AW4.7, GE Healthcare). The region of interest (ROI) was placed on the axial image showing the maximum area of the solid component, covering more than 70% of the lesion area and avoiding cavitation, necrosis, and major vessels. Corresponding ROIs were subsequently placed on the aorta of the same axial image for normalized iodine concentration (NIC) calculation (Figures 2,3). All measurements were repeated three times at three consecutive axial imaging levels to ensure consistency, and the average values were analyzed. Example of spectral parameters measurement in a solid nodule (a 67-year-old male, invasive adenocarcinoma). The ROI was placed on the axial image showing the maximum area of the solid component, covering more than 70% of the lesion area and avoiding cavitation, necrosis, and major vessels. Corresponding ROI was subsequently placed on the aorta at the same level for NIC calculation. “1” indicates the lesion ROI and “2” indicates the aorta ROI at the same axial level. (A) An axial conventional CT image with lung window settings (width 1,600 HU; level −650 HU); (B) effective atomic number map; (C) iodine concentration map. (D,E) Spectral HU attenuation curves showing the 40- and 100-keV CT values used to calculate λ. All image panels are de-identified. CT, computed tomography; HU, Hounsfield unit; NIC, normalized iodine concentration; ROI, region of interest. Example of spectral parameters measurement in a part-solid nodule (a 56-year-old male, invasive adenocarcinoma). The ROI was placed on the axial image showing the maximum area of the solid component, covering more than 70% of the lesion area and avoiding cavitation, necrosis, and major vessels. Corresponding ROI was subsequently placed on the aorta at the same level for NIC calculation. “1” indicates the lesion ROI and “2” indicates the aorta ROI at the same axial level. (A) An axial conventional CT image with lung window settings (width 1,600 HU; level –650 HU); (B) effective atomic number map; (C) iodine concentration map. (D,E) Spectral HU attenuation curves showing the 40- and 100-keV CT values used to calculate λ. All image panels are de-identified. CT, computed tomography; HU, Hounsfield unit; NIC, normalized iodine concentration; ROI, region of interest. The iodine concentration (IC) of lesions in the double-phase enhanced scan were measured in the iodine density image derived from the iodine/water-based material decomposition image. The IC in the descending aorta was also measured in the same slice. The NIC, which represents the ratio of IC in the lesion and descending aorta, was calculated. These IC parameters (i.e., IC and NIC) were calculated both in the arterial and venous phases. The slope of the spectral HU curve (λHU) was assessed only in the 40−100 keV region by the equation λHU = [(CT40 keV − CT100 keV) HU/60] based on previous studies. Each spectral CT parameter was independently measured by two chest radiologists (W.L. and W.C., with 2 and 3 years of experience, respectively) who were blinded to patients’ clinical data and histological diagnosis. Reproducibility of spectral CT quantitative parameters Inter-observer agreement was assessed using the initial parameters calculated from the two radiologists (W.L. and W.C.). Pathologic outcomes For each patient, the nodule’s final pathologic diagnosis (classified as STAS positive and negative) was extracted from the pathology reports which were reviewed and finalized by specialized chest pathologists with over 10 years’ experience. The diagnoses were determined on the bases of 5th edition of the WHO classification of lung tumors. STAS was defined as the presence of micropapillary clusters, solid nests, or single tumor cells spreading within air spaces in the surrounding lung parenchyma beyond the edge of the main tumor. Statistical analysis Statistical analyses were performed using SPSS 27.0 (IBM Corp., Armonk, NY, USA) and GraphPad Prism 9.5.1 (GraphPad Software, La Jolla, CA, USA). Continuous variables were summarized as means and standard deviations for normally distributed data or as medians and interquartile ranges for nonparametric data. Categorical variables were summarized as frequencies and percentages. The agreement of average spectral CT parameters measured by two radiologists was assessed using the intraclass correlation coefficient (ICC) with criteria of agreement categorized as poor (<0.50), fair (0.50–<0.75), moderate (0.75–<0.90), and excellent (≥0.90) (19). The normality of continuous variables was evaluated with the Shapiro-Wilk test. Comparisons of continuous variables between two groups were conducted using the independent samples t-test or the Mann-Whitney U test, as appropriate. Categorical variables were compared using the Chi-squared test or Fisher’s exact test. Variables identified as significant in univariate analysis were subsequently included in multivariate stepwise logistic regression to construct three predictive models: the conventional CT imaging feature model (conventional CT model), the spectral CT parameter model (spectral model), and the integrated model. The integrated model was constructed by integrating the spectral model and the conventional model. The discriminative performance of each model was evaluated using the AUC, and pairwise comparisons were performed using the DeLong test. A two-sided P value <0.05 was considered statistically significant. Results Demographic and clinical features Between 1 January 2025 and 10 November 2025, a total of 256 patients with suspected lung cancer who underwent dual-phase spectral CT scans prior to surgical resection at our institution were included. Among those, 92 patients were excluded because of pure GGNs (n=90) or poor CT image quality due to respiratory motion artifacts (n=2). Demographic and clinical characteristics are summarized in Table 1. A total of 164 patients were included, with 80 males (48.8%) and 84 females (51.2%). The mean age was 62.0±9.6 years. Among the 164 patients, 67 (40.9%) patients were STAS positive whereas 97 (59.1%) were STAS negative on pathological examination. There were no significant differences in sex, age, or smoking status between the STAS-positive and STAS-negative groups. Lobectomy was more frequent in the STAS-positive group than it was in the STAS-negative group (55.2% vs. 41.2%, P=0.049). Among serum tumor markers, CEA was significantly higher in the STAS-positive group (median 3.5 vs. 2.6 ng/mL, P=0.007), whereas no significant differences were observed for NSE, CYFRA21-1, and SCCA (all P>0.05). Table 1Demographic and clinical characteristics of all 164 patients based on STAS positive and negative statusVariableAll patients (n=164)STAS-positive group (n=67)STAS-negative group (n=97)t/Z/χ2 valueP valueSex2.8550.091 Male80 (48.8)38 (56.7)42 (43.3) Female84 (51.2)29 (43.3)55 (56.7)Age (years)62.0±9.662.2±7.561.8±11.2−0.2430.808Smoking status3.3720.185 Never120 (73.2)46 (68.7)74 (76.3) Former31 (18.9)17 (25.4)14 (14.4) Current13 (7.9)4 (6.0)9 (9.3)Surgical approach6.0240.049 Wedge resection28 (17.1)6 (9.0)22 (22.7) Segmentectomy59 (36.0)24 (35.8)35 (36.1) Lobectomy77 (47.0)37 (55.2)40 (41.2)NSE (ng/mL)13.1 (11.4–15.1)13.4 (11.0–15.4)12.9 (11.6–15.1)−0.5830.560CYFRA21-1 (ng/mL)2.8 (1.9–3.7)2.5 (1.8–3.2)3.2 (1.9–4.0)−1.0280.304CEA (ng/mL)2.9 (2.0–4.6)3.5 (2.1–6.8)2.6 (1.9–3.8)−2.6760.007SCCA (ng/mL)1.0 (0.7–1.7)1.0 (0.7–1.9)1.1 (0.7–1.7)−0.3430.732Data are presented as n (%), mean ± standard deviation, or median (Q1–Q3). Comparisons of continuous variables between two groups were conducted using the independent samples t-test or the Mann-Whitney U test; Categorical variables were compared using the Chi-squared test or Fisher’s exact test. CEA, carcinoembryonic antigen; CYFRA21-1, cytokeratin fragment 21-1; NSE, neuron-specific enolase; SCCA, squamous cell carcinoma antigen; STAS, spread through air spaces. Demographic and clinical characteristics of all 164 patients based on STAS positive and negative status Data are presented as n (%), mean ± standard deviation, or median (Q1–Q3). Comparisons of continuous variables between two groups were conducted using the independent samples t-test or the Mann-Whitney U test; Categorical variables were compared using the Chi-squared test or Fisher’s exact test. CEA, carcinoembryonic antigen; CYFRA21-1, cytokeratin fragment 21-1; NSE, neuron-specific enolase; SCCA, squamous cell carcinoma antigen; STAS, spread through air spaces. Pathological features The pathological characteristics of all 164 patients are summarized in Table 2. Patients with STAS positivity were more likely to present with advanced pathological stage, including higher T stage (T2–3: 25.4% vs. 6.2%, P=0.001), N stage (N1–2: 23.9% vs. 3.1%, P<0.001), and overall TNM stage (Ⅱ–Ⅲ: 28.4% vs. 5.2%, P<0.001). Lymphovascular invasion (32.8% vs. 5.2%), pleural invasion (26.9% vs. 6.2%), and lymph node metastasis (23.9% vs. 3.1%) were more frequently found in the STAS positive group (all P<0.001). Nerve invasion was not significantly associated with STAS status (P=0.066). Table 2Pathological features of all 164 patients based on STAS-positive and -negative statusVariableAll patients (n=164)STAS positive group (n=67)STAS negative group (n=97)χ2 valueP valueT stage†–0.001 1141 (86.0)50 (74.6)91 (93.8) 221 (12.8)16 (23.9)5 (5.2) 32 (1.2)1 (1.5)1 (1.0)N stage†–<0.001 0145 (88.4)51 (76.1)94 (96.9) 15 (3.0)4 (6.0)1 (1.0) 214 (8.5)12 (17.9)2 (2.1)TNM stage†–<0.001 I140 (85.4)48 (71.6)92 (94.8) II13 (7.9)8 (11.9)5 (5.2) III11 (6.7)11 (16.4)0 (0)Histological subtype9.7520.002 MIA13 (7.9)0 (0)13 (13.4) IAC151 (92.1)67 (100.0)84 (86.6)Lymphovascular invasion22.079<0.001 Present27 (16.5)22 (32.8)5 (5.2) Absent137 (83.5)45 (67.2)92 (94.8)Nerve invasion†–0.066 Present3 (1.8)3 (4.5)0 (0) Absent161 (98.2)64 (95.5)97 (100.0)Pleural invasion13.566<0.001 Present24 (14.6)18 (26.9)6 (6.2) Absent140 (85.4)49 (73.1)91 (93.8)Lymph node metastasis16.718<0.001 Present19 (11.6)16 (23.9)3 (3.1) Absent145 (88.4)51 (76.1)94 (96.9)T, N, and TNM stages were based on the 9th edition lung cancer stage classification. †, Fisher’s exact test. Categorical variables were compared using the Chi-squared test or Fisher’s exact test. IAC, invasive adenocarcinoma; MIA, minimally invasive adenocarcinoma; N, node; STAS, spread through air spaces; T tumor; TNM, tumor, node, metastasis. Pathological features of all 164 patients based on STAS-positive and -negative status T, N, and TNM stages were based on the 9th edition lung cancer stage classification. †, Fisher’s exact test. Categorical variables were compared using the Chi-squared test or Fisher’s exact test. IAC, invasive adenocarcinoma; MIA, minimally invasive adenocarcinoma; N, node; STAS, spread through air spaces; T tumor; TNM, tumor, node, metastasis. Conventional CT features Nodules with STAS positivity were more likely to be solid (82.1% vs. 34.0%, P<0.001), and exhibited higher frequencies of spiculation (85.1% vs. 57.7%, P<0.001), and lobulation (92.5% vs. 75.3%, P=0.006). The STAS-positive rate was 62.5% (55/88) in solid nodules and 15.8% (12/76) in part-solid nodules. STAS-positive nodules also had larger whole tumor size (median 21.00 vs. 16.00 mm, P<0.001), larger solid component size (median 19.50 vs. 8.50 mm, P<0.001), and higher CTR (median 1.00 vs. 0.56, P<0.001) than STAS-negative nodules. There were no significant differences between the STAS-positive and STAS-negative groups in features including nodule location, margin, shape, pleural retraction, bubble lucency, or air bronchogram (all P>0.05) (Table 3). Table 3Conventional CT features of all 164 patients based on STAS-positive and -negative statusVariableAll patients (n=164)STAS-positive group (n=67)STAS-negative group (n=97)Z/χ2 valueP valueNodule location (lobe)5.0250.289 Left upper42 (25.6)17 (25.4)25 (25.8) Left lower16 (9.8)5 (7.5)11 (11.3) Right upper61 (37.2)24 (35.8)37 (38.1) Right middle15 (9.1)4 (6.0)11 (11.3) Right lower30 (18.3)17 (25.4)13 (13.4)Consistency36.823<0.001 PSN76 (46.3)12 (17.9)64 (66.0) SN88 (53.7)55 (82.1)33 (34.0)Margin (%)1.8250.203 Clear74 (45.1)26 (38.8)48 (49.5) Unclear90 (54.9)41 (61.2)49 (50.5)Shape (%)0.5080.525 Regular74 (45.1)28 (41.8)46 (47.4) Irregular90 (54.9)39 (58.2)51 (52.6)Spiculation13.827<0.001 Present113 (68.9)57 (85.1)56 (57.7) Absent51 (31.1)10 (14.9)41 (42.3)Lobulation8.1290.006 Present135 (82.3)62 (92.5)73 (75.3) Absent29 (17.7)5 (7.5)24 (24.7)Pleural retraction0.0250.847 Present109 (66.5)45 (67.2)64 (66.0) Absent55 (33.5)22 (32.8)33 (34.0)Bubble lucency0.9670.358 Present40 (24.4)19 (28.4)21 (21.6) Absent124 (75.6)48 (71.6)76 (78.4)Air bronchogram1.6900.207 Present81 (49.4)29 (43.3)52 (53.6) Absent83 (50.6)38 (56.7)45 (46.4)Distance to pleura (mm)0.00 (0.00–9.00)0.00 (0.00–7.00)0.00 (0.00–11.50)−0.3820.702Whole tumor size (mm)17.50 (13.50–24.38)21.00 (16.00–27.50)16.00 (12.50–20.50)−4.192<0.001Solid component size (mm)11.75 (7.50–20.50)19.50 (13.50–26.50)8.50 (6.00–13.25)−6.979<0.001CTR0.85 (0.45–1.00)1.00 (0.91–1.00)0.56 (0.40–1.00)−6.145<0.001Data are presented as n (%) or median (Q1–Q3). Comparisons of continuous variables between two groups were conducted using the Mann-Whitney U test; Categorical variables were compared using the Chi-squared test or Fisher’s exact test. CT, computed tomography; CTR, consolidation/tumor ratio; PSN, part-solid nodule; SN, solid nodule; STAS, spread through air spaces. Conventional CT features of all 164 patients based on STAS-positive and -negative status Data are presented as n (%) or median (Q1–Q3). Comparisons of continuous variables between two groups were conducted using the Mann-Whitney U test; Categorical variables were compared using the Chi-squared test or Fisher’s exact test. CT, computed tomography; CTR, consolidation/tumor ratio; PSN, part-solid nodule; SN, solid nodule; STAS, spread through air spaces. Spectral CT parameters Table 4 presents the spectral CT parameters of all patients. The STAS-positive group showed significantly higher CT number at 100 keV in both arterial and venous phases (arterial phase: median 22.30 vs. −34.20 HU, P<0.001; venous phase: median 28.20 vs. −22.20 HU, P<0.001), along with lower λ in the arterial phase (median 2.10 vs. 2.72, P=0.003), and Zeff in both arterial and venous phases (arterial phase: median 8.70 vs. 9.09, P<0.001; venous phase: median 8.70 vs. 8.88, P=0.003). No significant differences were observed in λ in the venous phase, NIC in both arterial and venous phases, and NICr between STAS-positive and STAS-negative groups (all P>0.05). Table 4Spectral CT parameters of all 164 patients based on STAS-positive and -negative statusVariableAll patients (n=164)STAS positive group (n=67)STAS negative group (n=97)Z/χ2 valueP valueArterial phase CT number at 100 keV−0.05 (−97.40 to 25.90)22.30 (6.60 to 33.90)−34.20 (−208.70 to 2.30)−6.684<0.001 λ2.38 (1.77 to 3.15)2.10 (1.76 to 2.56)2.72 (1.78 to 3.37)−2.9340.003 Zeff8.82 (8.56 to 9.22)8.70 (8.52 to 8.85)9.09 (8.71 to 9.38)−4.528<0.001 NIC0.17 (0.12 to 0.22)0.16 (0.12 to 0.19)0.18 (0.12 to 0.24)−1.7680.077Venous phase CT number at 100 keV11.55 (−67.70 to 31.55)28.20 (14.50 to 37.10)−22.20 (−208.75 to 15.45)−6.329<0.001 λ2.23 (1.74 to 2.64)2.24 (1.82 to 2.54)2.17 (1.69 to 2.80)−0.6990.484 Zeff8.77 (8.52 to 9.02)8.70 (8.50 to 8.86)8.88 (8.56 to 9.06)−2.9470.003 NIC0.48 (0.38 to 0.62)0.47 (0.42 to 0.58)0.48 (0.36 to 0.65)−0.4370.662 NICr0.34 (0.28 to 0.42)0.33 (0.28 to 0.39)0.35 (0.27 to 0.45)−1.3260.185Data are presented as median (Q1 to Q3). λ energy attenuation curve slope; CT, computed tomography; NIC, normalized iodine concentration; NICr, NIC ratio; STAS, spread through air spaces; Zeff, effective atomic number. Spectral CT parameters of all 164 patients based on STAS-positive and -negative status Data are presented as median (Q1 to Q3). λ energy attenuation curve slope; CT, computed tomography; NIC, normalized iodine concentration; NICr, NIC ratio; STAS, spread through air spaces; Zeff, effective atomic number. Table S1 shows the inter-reader reliability for all spectral CT parameters. In the arterial phase, the inter-reader agreement was excellent in calculating λ and NIC with ICC values of 0.93 (95% CI: 0.86–0.97) and 0.94 (95% CI: 0.89–0.97) respectively. The inter-reader agreement was moderate for CT number at 100 keV and Zeff with ICCs of 0.87 (95% CI: 0.74–0.94) and 0.89 (95% CI: 0.78–0.94), respectively. In the venous phase, the inter-reader agreement was excellent in CT number at 100 keV (ICC =0.91, 95% CI: 0.81–0.96) and moderate for calculating λ, Zeff, and NIC with ICC values of 0.88 (95% CI: 0.76–0.94), 0.87 (95% CI: 0.75–0.94), and 0.81 (95% CI: 0.64–0.91), respectively. Development and performance of the prediction models Three predictive models were developed to differentiate STAS-positive from STAS-negative LUAD. Model 1 used conventional CT features in predicting STAS positivity. In multivariable logistic regression analysis, significant predictors of radiological features included lobulation [odds ratio (OR) =3.509, 95% CI: 1.059–11.627], whole tumor size (OR =1.060, 95% CI: 1.013–1.110), and CTR (OR =72.057, 95% CI: 14.619–355.176) (Table 5). The conventional CT model showed considerable predictive performance, with an AUC of 0.829 (95% CI: 0.763–0.883). Model 2 used spectral CT parameters in predicting STAS positivity. Significant predictors included arterial λ (OR =0.578, 95% CI: 0.396–0.846) and venous 100-keV CT number (OR =1.017, 95% CI: 1.009–1.024) (Table 5). The spectral CT model also showed considerable performance, with an AUC of 0.829 (95% CI: 0.762–0.883). The DeLong test showed no significant difference between these two AUCs (P=0.99). Table 5Univariate and multivariate analysis of conventional CT features and spectral CT parametersVariableUnivariateMultivariateOR95% CIP valueOR95% CIP valueNodule location (lobe)1.1290.901, 1.4150.292Consistency8.8894.188, 18.867<0.001Margin1.5450.821, 2.9070.178Shape1.2560.670, 2.3540.476Spiculation4.1731.906, 9.135<0.001Lobulation4.0771.468, 11.3200.0073.5091.059, 11.6270.040Pleural retraction1.0550.545, 2.0420.874Bubble lucency1.4330.699, 2.9380.327Air bronchogram0.6600.353, 1.2360.194Distance to pleura0.9980.958, 1.0400.916Whole tumor size1.0801.036, 1.126<0.0011.0601.013, 1.1100.011Solid component size1.1361.083, 1.190<0.001CTR80.16217.562, 365.913<0.00172.05714.619, 355.176<0.001Arterial phase CT number at 100 keV1.0161.009, 1.024<0.001 λ0.6310.456, 0.8710.0050.5780.396, 0.8460.005 Zeff0.2110.098, 0.455<0.001 NIC0.0270.000, 2.0120.101Venous phase CT number at 100 keV1.0171.009, 1.026<0.0011.0171.009, 1.024<0.001 λ0.8450.564, 1.2660.414 Zeff0.3110.124, 0.7780.013 NIC0.4770.081, 2.8010.412 NICr0.1090.007, 1.5800.104Integrated model CTR7.3921.124, 48.6190.037 λ†0.6130.420, 0.8940.011 CT number at 100 keV‡1.0121.004, 1.0200.003†, arterial phase; ‡, venous phase. λ, energy attenuation curve slope; CI, confidence interval; CT, computed tomography; CTR, consolidation/tumor ratio; NIC, normalized iodine concentration; NICr, NIC ratio; OR, odds ratio; Zeff, effective atomic number. Univariate and multivariate analysis of conventional CT features and spectral CT parameters †, arterial phase; ‡, venous phase. λ, energy attenuation curve slope; CI, confidence interval; CT, computed tomography; CTR, consolidation/tumor ratio; NIC, normalized iodine concentration; NICr, NIC ratio; OR, odds ratio; Zeff, effective atomic number. The integrated model combined conventional CT features with spectral CT parameters. In multivariable logistic regression analysis, significant predictors included CTR (OR =7.392, 95% CI: 1.124–48.619), arterial λ (OR =0.613, 95% CI: 0.420–0.894), and CT number at venous 100 keV (OR =1.012, 95% CI: 1.004–1.020) (Table 5). The integrated model achieved the numerically highest AUC of 0.842 (95% CI: 0.777–0.894) (Figure 4). However, pairwise DeLong tests showed no significant difference between the integrated model and the conventional CT model (P=0.51) or the spectral CT model (P=0.31), indicating that adding spectral CT parameters did not provide a clinically meaningful incremental discriminative value in this cohort. ROC curves for the conventional CT model, spectral model and integrated model combined CT features and spectral parameters. AUC, area under the curve; CI, confidence interval; CT, computed tomography; ROC, receiver operating characteristic. Discussion The presence of STAS has been demonstrated to be a significant predictor of increased postoperative recurrence and reduced survival, especially in patients with sublobar resection (3-5). In this study, all three radiological models showed considerable performance for preoperative STAS prediction in solid and part-solid LUAD nodules. Although the integrated model achieved the numerically highest AUC, the improvement over the conventional CT and spectral CT models was small and not statistically significant. This finding suggests that spectral CT parameters can provide comparable quantitative information but did not clearly improve discrimination beyond conventional CT features in this cohort. Among conventional CT features, lobulation, whole tumor size, and CTR were independently associated with STAS positivity. These findings are biologically plausible because larger solid components and lobulated margins may reflect more invasive growth patterns and a higher probability of tumor cell spread beyond the main tumor edge. In the present cohort, STAS positivity was substantially more frequent in solid nodules than it was in part-solid nodules, highlighting the importance of nodule consistency in preoperative risk assessment. Unlike several previous studies that included pure GGNs or broadly defined nonsolid nodules, our study focused on the solid and part-solid nodules most relevant to preoperative surgical decision-making and formally compared the incremental value of spectral CT against conventional CT features. The independent spectral CT predictors also have a plausible biological basis (18). STAS-positive tumors more frequently exhibit high-grade invasive histologic subtypes, such as micropapillary or solid components (20), which are characterized by increased cellular density and stromal fibrosis. A higher venous 100-keV CT value may therefore reflect increased soft-tissue components and delayed enhancement in more invasive tumors. Conversely, a lower arterial λ indicates a flatter spectral attenuation curve during early enhancement, potentially reflecting altered iodine distribution, vascular permeability, and microvascular perfusion associated with VEGF overexpression in STAS-positive tumors (21). Thus, arterial λ and venous 100-keV CT values may be indirect quantitative markers of tumor composition and perfusion-related aggressiveness rather than direct imaging signs of STAS. The lack of significant incremental value of the integrated model may be explained by information overlap between conventional CT features and spectral CT parameters. CTR, tumor size, and lobulation already capture morphologic aggressiveness and the extent of the solid invasive component, whereas spectral CT parameters partly reflect the same underlying differences in cellularity, stromal content, and enhancement. In addition, the modest sample size may have limited the statistical power to detect small incremental gains. Clinically, this means that conventional CT features remain a robust and readily available basis for STAS risk assessment, whereas spectral CT should currently be regarded as a quantitative adjunct rather than a replacement for or routine addition to conventional CT solely for STAS prediction. Compared with recent machine-learning approaches, our interpretable logistic-regression models achieved a similar range of discrimination while remaining easier to interpret clinically. A recent dual-center DLCT study by Fang et al. used machine learning classifiers incorporating DLCT parameters and clinicoradiological characteristics for STAS prediction (22). Such approaches may capture nonlinear interactions that are not modeled by standard logistic regression. Similarly, radiomics may further improve performance by quantifying intratumoral heterogeneity and tumor-lung interface patterns, as suggested by recent radiomics studies and meta-analyses (11,23). However, radiomics and machine learning models require robust segmentation, feature harmonization, and external validation to reduce overfitting and ensure transportability. From a clinical perspective, potential error patterns should also be considered. False-positive predictions may occur in STAS-negative nodules with a high CTR, lobulation, or high venous 100 keV attenuation, because these features can mimic invasive biological behavior without microscopic air-space spread. Conversely, false-negative predictions may occur in STAS-positive part-solid or small nodules with a limited solid component or less conspicuous enhancement, in which microscopic spread is not well reflected by macroscopic CT morphology or spectral attenuation. These patterns emphasize that model output should be integrated with surgical, pathological, and clinical judgment rather than used as a stand-alone decision rule. There were several limitations in this study. First, the retrospective single-center design may have introduced selection bias. Moreover, the prediction models were developed and evaluated in the same cohort without internal or external validation, which may have resulted in optimistic estimates of model performance. Therefore, further validation in independent multicenter cohorts is required. Second, STAS status was obtained from the original postoperative pathology reports rather than from a centralized retrospective review of all histologic slides, which may have introduced variability in pathological assessment. Third, this study used one fast kVp-switching spectral CT platform; differences in spectral separation, reconstruction algorithms, iodine quantification, and Zeff estimation across vendors may affect model transportability. Fourth, the sample size was relatively small, especially for STAS-positive part-solid nodules, which limited subgroup-specific model development. Fifth, body mass index (BMI) and objective image noise metrics were not available, so the influence of patient habitus on image noise, spiculation assessment, and model performance could not be directly evaluated. Sixth, Zeff values were generated using the vendor-provided software without local phantom-based calibration; therefore, Zeff-related findings should be interpreted in the context of vendor-specific reconstruction and published phantom accuracy data (24). Finally, photon-counting CT (PCCT) may improve spatial resolution, noise performance, and spectral quantification (25), but model coefficients derived from fast kVp-switching spectral CT cannot be directly transferred to PCCT systems without recalibration and validation. Conclusions Spectral CT parameters showed diagnostic performance comparable to that of conventional CT features for predicting STAS in solid and part-solid LUAD nodules. However, the integrated model did not provide a statistically significant incremental value over conventional CT in this cohort. Further multicenter, cross-vendor studies are warranted to validate these findings and to compare interpretable statistical models with radiomics and machine learning approaches. Supplementary The article’s supplementary files as