A combined model based on dual-energy CT for predicting high-risk lung adenocarcinoma in nodules with solid component

Abstract

Background

High-risk factors of lung adenocarcinoma (LUAD), including poorly differentiated pattern, lymphovascular invasion, visceral pleural invasion, lymph node metastasis, and spread through air spaces, are associated with poor prognosis.

Purpose

To construct and validate a combined model based on dual-energy computed tomography (DECT) for predicting high-risk LUAD in nodules with solid components.

Material and Methods

A total of 539 pathologically confirmed LUADs were retrospectively enrolled and were randomly divided into training (n = 377) and testing cohorts (n = 162) (ratio = 7:3). DECT quantitative parameters, conventional quantitative-semantic features, and clinical characteristics were collected. Univariable and multivariable logistic regression was used to construct the conventional CT model, DECT model, and combined model. Area under the curve (AUC) was used to evaluate diagnostic performance of models.

Results

The combined model (features: sex, diameter, pleural indentation, electron density in venous phase [Rho_V], and NIC_A [normalized iodine concentration in arterial phase]) showed significantly higher AUC than the DECT model (features: Rho_V and NIC_A) (0.899 vs. 0.870; P = 0.005) and the conventional CT model (features: diameter and pleural indentation) (0.899 vs. 0.742; P <0.001) in the training cohort. In the testing cohort, the combined model showed significantly higher AUC than the conventional CT model (0.852 vs. 0.704; P <0.001), but no significant difference in AUC was found between the combined model and the DECT model (0.852 vs. 0.838; P = 0.354).

Conclusion

The combined model based on DECT has good diagnostic performance in predicting high-risk LUAD in nodules with solid components.

Keywords

Dual-energy computed tomography lung adenocarcinoma high-risk factors quantitative-semantic features quantitative parameters

Introduction

Lung adenocarcinoma (LUAD) is the most prevalent histological type of lung cancer (1). High-risk factors of LUAD, including poorly differentiated pattern (2,3), lymphovascular invasion (LVI) (4), visceral pleural invasion (VPI) (5,6), lymph node (LN) metastasis (7), and spread through air spaces (STAS) (8), are associated with poor prognosis. Huang et al. suggested that segmentectomy lead to no difference in short-term oncologic outcomes between clinical stage IA3 and IA1-2 non-small cell lung cancer (NSCLC) (9). However, lobectomy and complete LN dissection could improve long-term prognosis compared with sublobar resection and incomplete LN dissection in small-sized LUAD (diameter ≤30 mm) with high-risk factors (10 –15). Besides, high-risk LUAD with VPI and LN metastasis (stage over IB) could benefit from neoadjuvant immunotherapy plus chemotherapy, resulting in longer event-free survival and a higher rate of pathological complete response (16). Therefore, preoperatively differentiating high-risk LUAD may be useful for nodule management and prognostic prediction.

Dual-energy computed tomography (DECT), which can obtain virtual non-contrast (VNC) image and multiple quantitative parameters, including iodine concentration (IC), electron density (Rho), slope of the spectral curve (λ_HU), and effective atomic number (Zeff), is commonly used for diagnosing pulmonary nodules. The VNC image may have the potential to replace the true non-contrast (TNC) image in chest imaging for reducing radiation dose (17,18). Among the DECT quantitative parameters, the normalized IC (NIC) could differentiate malignant solid nodules (19,20), predict LVI in NSCLC (21), and distinguish STAS in solid LUAD (22). The Rho could predict the LUAD invasiveness in ground-glass nodules (GGNs) (23,24). The λHU could predict high-grade NSCLC (25) and solid/micropapillary predominant LUAD (26). In addition, conventional CT quantitative-semantic features could predict high-risk LUAD in part-solid and solid nodules (27) and differentiate the invasiveness of LUAD in GGNs (28). However, to our knowledge, no previous studies predicted high-risk LUAD by integrating DECT quantitative parameters and conventional CT quantitative-semantic features.

Pure GGNs rarely have high-risk factors because of their indolent nature (29 –33). Therefore, the aim of the present study was to construct and validate a combined model based on DECT integrating DECT quantitative parameters and conventional CT quantitative-semantic features in predicting high-risk LUAD in nodules with solid components.

Material and Methods

Patients

The Ethics Committee of Sichuan Cancer Hospital approved the retrospective study (ethical number: SCCHEC-02-2021-088, approval year: 2021) and the requirement for informed consent was waived. A total of 802 patients who underwent contrast-enhanced DECT between April 2021 and April 2025 were retrospectively enrolled. Inclusion criteria included the following: (i) pulmonary nodules (diameter ≤3 cm) with complete surgical resection; (ii) pathological diagnosis with LUAD; (iii) DECT examination within 14 days before surgery; and (iv) complete pathological data. The exclusion criteria were as follows: (i) pure GGNs; (ii) received anti-cancer treatment before DECT; (iii) poor image quality; and (iv) a history of other malignant tumors. The nodule with the largest diameter was selected for analysis when the patient had several resected nodules.

Finally, our study included 539 patients. All nodules were randomly divided into training (n = 377) and testing cohorts (n = 162) at a 7:3 ratio (Fig. 1).

Fig. 1.

The flow chart of nodules selection. DECT, dual-energy computed tomography.

Image acquisition

All contrast-enhanced DECT examinations were carried out with dual-source CT (Somatom Definition Flash; Siemens Healthcare, Forchheim, Germany). A contrast agent of 80–90 mL (370 mg I/mL) was administered through the median cubital vein (2.5–3.0 mL/s), followed by 30–40 mL of saline (2.5–3.0 mL/s). All contrast-enhanced DECT of the chest was performed using the bolus-tracking method. When the threshold of the ascending aorta at the pulmonary trunk layer reached 100 HU, the arterial phase was initiated with a delay of 5 s. Then, the venous phase was initiated 30 s after the arterial phase.

The scanning parameters were as follows: reference current = 205/87 mAs; tube voltage = 80/Sn140 kV; pitch = 0.55; collimation = 64 × 0.6 mm; and rotation time = 0.28 s. The automated tube current modulation (CARE Dose 4D) was activated in this study.

Image reconstruction

The VNC image, which was used to assess the conventional CT quantitative-semantic features, was reconstructed using DECT data from the arterial phase, while the virtual monoenergetic image series (40 and 100 keV) were reconstructed using DECT data from the arterial and venous phases on a commercial workstation (SyngoVia VB40; Siemens Healthcare). The reconstruction parameters were as follows: iterative algorithm = SAFIRE (strength level 4); matrix = 512 × 512; reconstruction kernel = Q30f; field of view = 350 × 350 mm; slice thickness = 0.5 mm; and slice increment = 0.5 mm.

Conventional CT quantitative features

All nodules were automatically detected and segmented using an artificial intelligence (AI) platform (United Imaging Healthcare) based on deep learning algorithms (34). All segmentation results were also evaluated by two thoracic radiologists (JL and HQ, with 7 and 12 years of experience, respectively) in the lung window (level = −500 HU; width = 1500 HU). No manual adjustments were conducted due to the excellent segmentation results, even in the nodules with complex margins and adjacent vessels. The criteria for satisfaction were as follows: the segmentation included complete pulmonary nodules, while avoiding vessels and adjacent structures.

The AI-derived quantitative features of nodules, including axial maximal diameter, axial perpendicular diameter, mean attenuation, and volume, were recorded. The diameter was calculated with the formula: diameter = (axial maximal diameter + axial perpendicular diameter)/2. The axial maximal diameter of the solid component was manually measured by the aforementioned two radiologists. The consolidation:tumor ratio (CTR) was calculated with the formula: CTR = axial maximal diameter of the solid component/axial maximal diameter of the nodule.

To evaluate the performance of AI in automatically segmenting nodules, a senior thoracic radiologist (PZ, with 27 years of experience) selected 75 nodules for manual segmentation. Then, the inter-observer reproducibility of diameter, mean attenuation, and volume assessed by AI and the senior radiologist was evaluated.

Conventional CT semantic features

Two thoracic radiologists (JL and HQ, with 7 and 12 years of experience, respectively) who were blinded to histopathological results assessed the following semantic features: nodule type, spiculation, lobulation, air bronchogram, pleural indentation, pseudocavity, and vascular convergence sign (35). When the assessment results of the two radiologists were inconsistent, a senior thoracic radiologist (PZ, with 27 years of experience) resolved the disagreement.

DECT quantitative parameters

Two thoracic radiologists (JL and HQ, with 7 and 12 years of experience, respectively) placed circular regions of interest (ROIs) on the nodule and the aorta on the axial slice, which shows the maximum diameter of nodule. The circular ROIs were drawn as large as feasible while avoiding vessels, adjacent structures, and artifacts. The Zeff, Rho, IC (IC_nodule), attenuation of 40 keV (Attenuation_40keV), and attenuation of 100 keV (Attenuation_100keV) of the nodule and IC of the aorta (IC_aorta) were recorded. Then the λ_HU and NIC were respectively calculated with the formulas (25): λ_HU = (Attenuation_40keV–Attenuation_100keV)/60, NIC = IC_nodule/IC_aorta. Finally, a total of eight DECT quantitative parameters, including Zeff, Rho, NIC, and λ_HU in the arterial phase (Zeff_A, Rho_A, NIC_A, and λ_HU_A) and the venous phase (Zeff_V, Rho_V, NIC_V, and λ_HU_V), were recorded. The representative images were presented in Fig. 2.

Fig. 2.

The representative DECT images of nodules. (a–e) A 55-year-old man with low-risk LUAD manifesting as part-solid nodule without pleural indentation. (f–j) A 44-year-old man with high-risk LUAD manifesting as solid nodule with pleural indentation. The high-risk LUAD showed higher diameter, Rho_A, Rho_V, Zeff_A, and Zeff_V, and lower NIC_A and NIC_V than the low-risk LUAD. A, arterial phase; DECT, dual-energy computed tomography; LUAD, lung adenocarcinoma; NIC, normalized iodine concentration; Rho, electron density; V, venous phase; Zeff, effective atomic number.

Reproducibility of manual quantitative parameters

A total of 75 nodules in this study were randomly selected for the reproducibility analysis. For inter-observer reproducibility, the CTR and DECT quantitative parameters were independently measured by two radiologists (JL and HQ, with 7 and 12 years of experience, respectively). For intra-observer reproducibility, the CTR and DECT quantitative parameters were measured again by a radiologist (JL) after 1 month.

Histopathological evaluation

All pathological diagnoses were obtained based on the classification of lung tumors (36). The high-risk factors in invasive adenocarcinomas were as follows: (i) poorly differentiated pattern (grade 3), which has ≥20% high-grade pattern, including micropapillary, solid, cribriform, and complex glandular patterns; (ii) LVI; (iii) VPI; (iv) LN metastasis; and (v) STAS. All LUADs were divided into a low-risk group (including invasive adenocarcinomas without high-risk factors and minimally invasive adenocarcinomas) and a high-risk group (including invasive adenocarcinomas with any of the high-risk factors) (Table 1).

Table 1.

Demographic and conventional CT features of nodules.

Characteristic	Training cohort			P*	Testing cohort			P ^†
Characteristic	All (n = 377)	Low risk (n = 228)	High risk (n = 149)	P*	All (n = 162)	Low risk (n = 100)	High risk (n = 62)	P ^†
Sex				<0.001				0.175
Male	126	56	70		64	31	33
Female	251	172	79		98	69	29
Age (years)	58.0 (52.0–66.0)	57.0 (50.0–66.0)	60.0 (54.5–67.0)	0.002	59.0 (52.0–68.0)	57.0 (50.0–68.0)	61.0 (54.0–68.3)	0.488
BMI (kg/m²)	23.3 (21.1–25.0)	23.1 (21.0–24.9)	23.6 (21.7–25.3)	0.161	23.4 (21.5–25.3)	23.5 (21.2–25.7)	23.3 (21.8–24.6)	0.386
Diameter (mm)	15.2 (11.0–20.3)	13.2 (9.6–18.2)	18.5 (13.7–23.0)	<0.001	15.7 (11.9–20.1)	14.4 (11.2–18.2)	17.2 (12.5–22.7)	0.748
Attenuation (HU)	−329.5 (−510.2–−104.5)	−464.4 (−560.2–−274.2)	−136.5 (−258.6–−6.0)	<0.001	−316.8 (−536.6–−97.3)	−422.5 (−584.9–−218.5)	−137.3 (−296.2–10.8)	0.964
CTR (%)	71.2 (44.8–100.0)	53.4 (37.3–82.2)	100.0 (76.4–100.0)	<0.001	73.5 (42.5–100.0)	50.7 (38.0–100.0)	100.0 (97.3–100.0)	0.647
Volume (cm³)	1.8 (0.7–3.9)	1.2 (0.4–2.6)	2.8 (1.3–5.8)	<0.001	1.7 (0.8–3.7)	1.4 (0.7–2.5)	2.6 (1.1–5.5)	0.852
Nodule type				<0.001				0.355
Part-solid	221	175	46		88	73	15
Solid	156	53	103		74	27	47
Spiculation				<0.001				0.477
No	280	193	87		125	86	39
Yes	97	35	62		37	14	23
Lobulation				0.001				0.060
No	85	64	21		25	18	7
Yes	292	164	128		137	82	55
Air bronchogram				0.447				0.408
No	254	157	97		115	69	46
Yes	123	71	52		47	31	16
Pleural indentation				<0.001				0.854
No	110	95	15		46	37	9
Yes	267	133	134		116	63	53
Pseudocavity				0.072				0.239
No	326	203	123		146	91	55
Yes	51	25	26		16	9	7
Vascular convergence				0.001				0.309
No	143	102	41		54	44	10
Yes	234	126	108		108	56	52
Pathological subtypes
MIA	24	24	0		11	11	0
Invasive adenocarcinomas	353	204	149		151	89	62
High-risk factors
Grade 3	41	0	41		18	0	18
VPI	69	0	69		26	0	26
LVI	40	0	40		16	0	16
STAS	89	0	89		37	0	37
LNM	28	0	28		15	0	15

Values are given as n or median (IQR).

*Group comparison between low-risk group and high-risk group in training cohort.

†

Group comparison between training cohort and testing cohort.

BMI, body mass index; CTR, consolidation:tumor ratio; LNM, lymph node metastasis; LVI, lymphovascular invasion; MIA, minimally invasive adenocarcinomas; STAS, spread through air space; VPI, visceral pleural invasion.

Statistical analysis

This study used R (version 4.4.1) and MedCalc (version 18.2.1) for statistical analysis. The continuous and categorical variables were compared using the Mann–Whitney U-test and Fisher's exact test, respectively. The reproducibility of CTR and DECT quantitative parameters was assessed using the intraclass correlation coefficient (ICC) and Bland–Altman plot. The univariable and multivariable logistic regression (with the backward stepwise selection) was used to construct the conventional CT model, DECT model, and combined model in the training cohort. Then, the testing cohort was used to validate these models. The diagnostic performance was assessed with the area under the curve (AUC). The AUC of models was compared with the DeLong test (37). A two-tailed P <0.05 showed a statistical difference.

Results

Clinical and conventional CT quantitative-semantic features of nodules

In the training cohort, the high-risk group demonstrated significantly higher age, diameter, attenuation, CTR, and volume than the low-risk group (all P <0.05). There were significant differences in sex, nodule type, spiculation, lobulation, vascular convergence sign, and pleural indentation between the two groups (all P <0.05), except for body mass index (BMI) (P = 0.161), air bronchogram (P = 0.447), and pseudocavity (P = 0.072) (Table 1).

No significant differences in sex, BMI, age, diameter, attenuation, CTR, volume, nodule type, spiculation, lobulation, air bronchogram, pleural indentation, pseudocavity, and vascular convergence sign were found between the training and testing cohorts (all P >0.05) (Table 1).

The ICCs of diameter, attenuation, and volume assessed by AI and the senior radiologist were 0.916, 0.987, and 0.933, respectively, indicating a satisfactory reproducibility.

Reproducibility analysis of manual quantitative parameters

The CTR and quantitative parameters of DECT showed good inter-observer agreement (ICC = 0.809–0.935) and intra-observer agreement (ICC = 0.811–0.956) in this study (Table S1). The Bland–Altman plots were provided in Figures S1 and S2 in the supplementary material.

DECT quantitative parameters of nodules

In the training cohort, the high-risk group showed significantly higher Rho_A, Rho_V, Zeff_A, and Zeff_V and significantly lower NIC_A, λ_HU_A, and λ_HU_V than the low-risk group (all P <0.05). However, there was no significant difference in NIC_V between the two groups (P = 0.543) (Table 2).

Table 2.

DECT quantitative parameters of nodules.

Characteristic	Training cohort			P*	Testing cohort			P ^†
Characteristic	All (n = 377)	Low risk (n = 228)	High risk (n = 149)	P*	All (n = 162)	Low risk (n = 100)	High risk (n = 62)	P ^†
Rho_A (HU)	−136.7 (−349.4–29.4)	−291.8 (−430.6–−76.9)	29.6 (−95.0–41.6)	<0.001	−98.3 (−363.4–33.8)	−271.3 (−411.0–−51.2)	31.0 (−30.0–43.0)	0.588
Rho_V (HU)	−140.9 (−342.6–31.2)	−279.2 (−400.1–−83.0)	33.3 (−88.0–43.3)	<0.001	−119.5 (−346.2–30.9)	−276.9 (−424.1–−43.5)	28.7 (−54.4–48.0)	0.656
NIC_A (%)	11.9 (7.5–16.7)	14.0 (9.4–19.1)	9.2 (5.3–13.0)	<0.001	12.8 (8.1–18.6)	14.9 (9.5–19.5)	10.0 (5.8–15.0)	0.164
NIC_V (%)	31.7 (24.5–42.6)	32.4 (23.7–43.3)	31.1 (25.0–41.3)	0.543	32.7 (23.5–41.6)	33.3 (22.4–43.3)	31.3 (23.9–40.8)	0.873
Zeff_A	7.8 (6.8–8.2)	7.4 (6.3–8.2)	8.0 (7.6–8.2)	<0.001	7.9 (6.8–8.3)	7.4 (6.0–8.2)	8.0 (7.8–8.3)	0.556
Zeff_V	8.1 (7.2–8.6)	7.7 (6.5–8.5)	8.4 (8.0–8.7)	<0.001	8.2 (7.1–8.6)	7.7 (6.3–8.6)	8.3 (8.1–8.6)	0.978
λ_HU_A	1.9 (1.3–2.6)	2.2 (1.7–2.8)	1.4 (0.9– 2.0)	<0.001	1.9 (1.4–2.7)	2.2 (1.6–2.9)	1.5 (1.0–2.1)	0.626
λ_HU_V	2.5 (1.9–3.1)	2.7 (2.0–3.2)	2.4 (1.7–2.9)	0.003	2.5 (1.9–3.0)	2.6 (2.1–3.1)	2.3 (1.7–2.8)	0.588

Values are given as median (IQR).

*Group comparison between low-risk group and high-risk group in training cohort.

†

Group comparison between training cohort and testing cohort.

λ_HU, slope of the spectral curve; A, arterial phase; DECT, dual-energy computed tomography; NIC, normalized iodine concentration; Rho, electron density; V, venous phase; Zeff, effective atomic number.

No significant differences in Rho_A, Rho_V, Zeff_A, Zeff_V, NIC_A, NIC_V, λ_HU_A, and λ_HU_V were found between the training cohort and testing cohort (all P >0.05) (Table 2).

Conventional CT model

Diameter, attenuation, CTR, volume, nodule type, spiculation, lobulation, pleural indentation, and vascular convergence sign were predictive factors for high-risk LUAD in the univariate logistic regression (all P <0.05). Multivariable logistic regression found that diameter and pleural indentation were the independent factors for high-risk LUAD (all P <0.05), which were used to construct the conventional CT model (Table 3). The calculation formula for the conventional CT model was shown in Table S2 in the supplementary material.

Table 3.

Univariable and multivariable logistic regression analysis for predicting high-risk lung adenocarcinomas in training cohort.

	Univariable analysis		Multivariable analysis
Models	Odds ratio	P	Odds ratio	P
Conventional CT model
Diameter (mm)	1.122 (1.082–1.165)	<0.001	1.089 (1.047–1.132)	<0.001
Attenuation (HU)	1.005 (1.004–1.007)	<0.001
CTR (%)	1.042 (1.032–1.051)	<0.001
Volume (cm³)	1.266 (1.166–1.374)	<0.001
Nodule type
Part-solid	Reference
Solid	7.393 (4.649–11.758)	<0.001
Spiculation
No	Reference
Yes	3.930 (2.418–6.386)	<0.001
Lobulation
No	Reference
Yes	2.379 (1.380–4.100)	0.002
Pleural indentation
No	Reference
Yes	6.381 (3.519–11.569)	<0.001	4.403 (2.367–8.190)	<0.001
Vascular convergence
No	Reference
Yes	2.132 (1.367–3.325)	0.001
DECT model
Rho_A (HU)	1.007 (1.005–1.008)	<0.001
Rho_V (HU)	1.007 (1.005–1.008)	<0.001	1.007 (1.005–1.008)	<0.001
NIC_A (%)	0.860 (0.824–0.897)	<0.001	0.856 (0.816–0.898)	<0.001
Zeff_A	1.738 (1.399–2.158)	<0.001
Zeff_V	2.123 (1.665–2.707)	<0.001
λ_HU_A	0.346 (0.258–0.464)	<0.001
λ_HU_V	0.667 (0.521–0.854)	0.001
Combined model
Sex
Male	Reference
Female	0.367 (0.236–0.571)	<0.001	0.412 (0.226–0.751)	0.004
Age (years)	1.037 (1.015–1.058)	0.001
Diameter (mm)	1.122 (1.082–1.165)	<0.001	1.097 (1.044–1.153)	<0.001
Pleural indentation
No	Reference
Yes	6.381 (3.519–11.569)	<0.001	2.125 (1.007–4.486)	0.048
Rho_V (HU)	1.007 (1.005–1.008)	<0.001	1.006 (1.004–1.008)	<0.001
NIC_A (%)	0.860 (0.824–0.897)	<0.001	0.859 (0.817–0.904)	<0.001

λ_HU, slope of the spectral curve; A, arterial phase; CTR, consolidation:tumor ratio; NIC, normalized iodine concentration; Rho, electron density; V, venous phase; Zeff, effective atomic number.

DECT model

Rho_A, Rho_V, NIC_A, Zeff_A, Zeff_V, λ_HU_A, and λ_HU_V were predictive factors for high-risk LUAD in the univariate logistic regression (all P <0.05). Multivariable logistic regression found that Rho_V and NIC_A were the independent factors for high-risk LUAD (all P <0.05), which were used to construct the DECT model (Table 3). The calculation formula for the DECT model was shown in Table S2 in the supplementary material.

Combined model

Furthermore, the combined model was constructed using multivariable logistic regression by integrating the clinical features with significant differences (sex and age) and all independent factors from the conventional CT model (diameter and pleural indentation) and DECT model (Rho_V and NIC_A). We found that sex, diameter, pleural indentation, Rho_V, and NIC_A were the independent factors for high-risk LUAD (all P <0.05), which were used to construct the combined model (Table 3). The calculation formula for the combined model was shown in Table S2 in the supplementary material.

Diagnostic performance

In the training cohort, the combined model showed significantly higher AUC than the DECT model (0.899 vs. 0.870, Z = 2.817; P = 0.005) and the conventional CT model (0.899 vs. 0.742, Z = 6.606; P <0.001). In the testing cohort, the combined model showed significantly higher AUC than the conventional CT model (0.852 vs. 0.704, Z = 3.598; P <0.001), but no significant difference in AUC was found between the combined model and the DECT model (0.852 vs. 0.838, Z = 0.927; P = 0.354).

In addition, the DECT model demonstrated significantly higher AUC than the conventional CT model in both the training (0.870 vs. 0.742, Z = 4.209; P <0.001) and testing cohorts (0.838 vs. 0.704, Z = 2.737; P = 0.006) (Table 4 and Fig. 3).

Fig. 3.

ROC curves of models in predicting high-risk lung adenocarcinoma in nodules with solid component in (a) the training cohort and (b) the testing cohort. AUC, area under the ROC curve; DECT, dual-energy computed tomography; ROC, receiver operating characteristic.

Table 4.

Diagnostic performance of models.

Models	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV	NPV	Cutoff	P ^*	P ^†
Training cohort
DECT model	0.870 (0.832–0.902)	0.725 (0.646–0.795)	0.882 (0.832–0.920)	0.800	0.831	0.550	0.005	−
Conventional CT model	0.742 (0.695–0.786)	0.785 (0.711–0.848)	0.627 (0.561–0.690)	0.579	0.817	0.393	<0.001	<0.001
Combined model	0.899 (0.864–0.927)	0.832 (0.762–0.888)	0.833 (0.778–0.879)	0.761	0.883	0.422	−	0.005
Testing cohort
DECT model	0.838 (0.772–0.891)	0.758 (0.633–0.858)	0.828 (0.739–0.897)	0.723	0.844	0.438	0.354	−
Conventional CT model	0.704 (0.627–0.773)	0.774 (0.650–0.871)	0.596 (0.493–0.693)	0.545	0.808	0.391	<0.001	0.006
Combined model	0.852 (0.787–0.903)	0.790 (0.668–0.883)	0.828 (0.739–0.897)	0.742	0.863	0.227	−	0.354

AUC comparison with combined model.

†

AUC comparison with DECT model.

AUC, area under the curve; CI, confidence interval; DECT, dual-energy computed tomography; NPV, negative predictive value; PPV, positive predictive value.

Discussion

Our study constructed the combined model (features: sex, diameter, pleural indentation, Rho_V, and NIC_A), the conventional CT model (features: diameter and pleural indentation), and the DECT model (features: Rho_V and NIC_A) to predict high-risk LUAD in nodules with solid components. We demonstrated that the combined model showed good performance in the training and testing cohorts, suggesting the potential of the combined model based on DECT in predicting high-risk LUAD as a non-invasive biomarker.

The high-risk factors of LUAD indicate the biological invasiveness. The VPI was identified as a non–size-dependent T2 identification and a prognostic predictor in TNM staging of LUAD, and the STAS has been identified as an invasion mode of lung cancer (36). The LUAD with high-risk factors has a poor prognosis (38 –42). Therefore, accurately predicting high-risk LUAD during preoperative DECT could facilitate the nodule management. The high-risk factors can be diagnosed through preoperative and intraoperative pathology. However, preoperative biopsy and intraoperative frozen sections have inadequate accuracy in identifying high-risk factors because of sampling issues (43,44). The DECT quantitative parameters have an advantage in assessing the benign or malignant nature of solid nodules (19) and predicting the invasiveness of LUAD in GGNs (28). Therefore, they may have the potential to predict high-risk LUAD in nodules with solid component, and this study is intended to establish a DECT model as a decision-support tool to alert surgeons to high-risk features that might be missed by limited sampling.

Rho represents the likelihood of an electron occurring at a certain position (45) and is positively correlated with tissue density (46). The high-grade tumor may have higher tumor density due to the high degree of malignancy (47). IC reveals the actual enhancement value and reflects underlying microvessel density and tumor angiogenesis (48,49). The high-grade tumor may cause necrosis in the central region of the tumor due to faster growth speed (50), which may result in a decrease in NIC. Therefore, our study demonstrated that the higher Rho_A and lower NIC_V were associated with high-risk LUAD in the DECT model. Previous studies demonstrated that Rho could distinguish high-grade from low-grade cerebral gliomas (51) and differentiate invasive from minimally invasive adenocarcinoma in pure GGNs (24), and NIC could differentiate invasive from preinvasive adenocarcinomas (52), which was similar to our study.

λ_HU, offering attenuation changes of material at different energy levels (53) could differentiate high-grade from low-grade NSCLC in solid nodules (25). Zeff, reflecting the composite atom of a mixture of distinct materials (54), could differentiate invasive from non-invasive LUAD in part-solid nodules (55). However, λHU and Zeff were not the independent factors in differentiating high-risk LUAD in this study. This difference might be related to the different types of included pulmonary nodules. Previous research focused on solid nodules or part-solid nodules, whereas our study evaluated part-solid and solid nodules.

In quantitative-semantic features, a larger tumor diameter suggests a greater likelihood of high-grade pattern in LUAD (56). Pleural indentation presents scar contraction, which was caused by the fibrotic hyperplasia (57). In addition, active fibroblast proliferation indicates the invasive growth pattern of the tumor (58). Therefore, diameter and pleural indentation were the independent factors in differentiating high-risk LUAD in our conventional CT model. Previous studies demonstrated that diameter could predict poorly differentiated from well/moderately differentiated LUAD (59), and pleural indentation could differentiate acinar from lepidic predominant LUAD (56), which was similar to our study.

In terms of diagnostic performance, the combined model and the DECT model demonstrated higher AUC than the conventional CT model in the training cohort and testing cohort. The combined model showed significantly higher AUC than the DECT model in the training cohort, but no significant difference in AUC was found between the combined model and the DECT model in the testing cohort. Therefore, both the combined model and the DECT model may be helpful in predicting the high-risk LUAD in nodules with solid components. Although the AUC of the combined model and the DECT model (ranging = 0.838–0.899) is statistically significant in this study, it falls short of the clinical threshold (typically >0.90) required to alter high-stakes surgical management, such as deciding between a lobectomy and a segmentectomy.

In clinical practice, when a pulmonary nodule with solid component is detected, we may predict high-risk LUAD by integrating sex, diameter, pleural indentation, Rho_V, and NIC_A with the calculation formula of the combined model.

The present study has some limitations. First, this was a single-center study. Although we have validated the combined model using an internal testing cohort. However, further validation is needed using external testing cohort. Second, potential selection bias may not be avoided in this retrospective study. Third, the quantitative-semantic features were assessed on VNC images. A previous study demonstrated that the VNC image may replace the TNC image in the measurement of attenuation and volumetry in the lung tumor model (17). Finally, we only used the axial slice to measure the maximal and perpendicular diameter, which might lead to under-staging.

In conclusion, the combined model based on DECT integrating sex, diameter, pleural indentation, Rho_V, and NIC_A has good diagnostic performance in predicting high-risk LUAD in nodules with solid components, which may be useful for nodule management and prognostic prediction.

Supplemental Material

sj-docx-1-acr-10.1177_02841851261452150 - Supplemental material for A combined model based on dual-energy CT for predicting high-risk lung adenocarcinoma in nodules with solid component

Supplemental material, sj-docx-1-acr-10.1177_02841851261452150 for A combined model based on dual-energy CT for predicting high-risk lung adenocarcinoma in nodules with solid component by Changjiu He, Xianyu Cao, Haomiao Qing, Peng Zhou, Yi Zhang, Xiaolei Dong and Jieke Liu in Acta Radiologica

Footnotes

ORCID iDs

Peng Zhou

Jieke Liu

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Sichuan Provincial Natural Science Foundation, National Natural Science Foundation of China (grant nos. 2025ZNSFSC1765, 2023NSFSC1635, 82202141).

Supplementary material

Supplementary material for this article is available online.

References

National Lung Screening Trial Research T. Lung cancer incidence and mortality with extended follow-up in the national lung screening trial. J Thorac Oncol 2019;14:1732–1742.

Yanagawa

Sugai

Shikanai

, et al. The new IASLC grading system for invasive non-mucinous lung adenocarcinoma is a more useful indicator of patient survival compared with previous grading systems. J Surg Oncol 2023;127:174–182.

Hou

Wang

Chen

, et al. Prognostic and predictive value of the newly proposed grading system of invasive pulmonary adenocarcinoma in Chinese patients: a retrospective multicohort study. Mod Pathol 2022;35:749–756.

Locke

Aung

Esposito

, et al. Lymphovascular invasion is a predictor of postoperative recurrence or death in stage I non-small-cell lung cancer (NSCLC). Clin Lung Cancer 2025;26:e560–e565.

Altorki

Wang

Damman

, et al. Recurrence of non-small cell lung cancer with visceral pleural invasion: a secondary analysis of a randomized clinical trial. JAMA Oncol 2024;10:1179–1186.

Akyil

Bayram

. Middle lobe tumors and lymphovascular invasion as independent predictors of recurrence-free survival in stage I NSCLC. BMC Pulm Med 2025;25:93.

Cai

Yang

Wang

. Occult lymph node metastasis is not a favorable factor for resected NSCLC patients. BMC Cancer 2023;23:822.

Shimomura

Miyagawa-Hayashino

Omatsu

, et al. Spread through air spaces is a powerful prognostic predictor in patients with completely resected pathological stage I lung adenocarcinoma. Lung Cancer 2022;174:165–171.

Huang

Brunelli

Stefanou

, et al. Is segmentectomy potentially adequate for clinical stage IA3 non-small cell lung cancer. Interdiscip Cardiovasc Thorac Surg 2025;40:ivaf064.

10.

Akamine

Wakasu

Matsubara

, et al.

Is sublobar resection feasible for high-risk pathologic stage I non-small cell lung cancer?

Ann Surg Oncol 2025;32:4161–4172.

11.

Sullivan

Rieger-Christ

, et al. Vascular invasion predicts the subgroup of lung adenocarcinomas ≤ 2.0 cm at risk of poor outcome treated by wedge resection compared to lobectomy. JTCVS Open 2023;16:938–947.

12.

Jeon

Lee

Shin

, et al. Prognostic impact of micropapillary and solid histological subtype on patients undergoing curative resection for stage I lung adenocarcinoma according to the extent of pulmonary resection and lymph node assessment. Lung Cancer 2022;168:21–29.

13.

Eguchi

Kameda

, et al. Lobectomy is associated with better outcomes than sublobar resection in spread through air spaces (STAS)-positive T1 lung adenocarcinoma: a propensity score-matched analysis. J Thorac Oncol 2019;14:87–98.

14.

Huang

Deng

Lin

, et al. Treatment modality for stage IB peripheral non-small cell lung cancer with visceral pleural invasion and ≤3 cm in size. Front Oncol 2022;12:830470.

15.

Yun

Lee

Choi

, et al. Comparison of prognostic impact of lymphovascular invasion in stage IA non-small cell lung cancer after lobectomy versus sublobar resection: a propensity score-matched analysis. Lung Cancer 2020;146:105–111.

16.

Forde

Spicer

, et al. Neoadjuvant nivolumab plus chemotherapy in resectable lung cancer. N Engl J Med 2022;386:1973–1985.

17.

Hering

Kroger

Bauer

, et al. Comparison of virtual non-contrast dual-energy CT and a true non-contrast CT for contouring in radiotherapy of 3D printed lung tumour models in motion: a phantom study. Br J Radiol 2020;93:20200152.

18.

Strotzer

Schachner

Scheuermeyer

, et al. Quantitative chest computed tomography: regional differences in dual-energy-derived virtual vs. true non-contrast scans. Clin Radiol 2025;85:106912.

19.

Liu

, et al. Quantitative parameters of enhanced dual-energy computed tomography for differentiating lung cancers from benign lesions in solid pulmonary nodules. Front Oncol 2022;12:1027985.

20.

Gonzalez-Perez

Arana

Barrios

, et al. Differentiation of benign and malignant lung lesions: dual-Energy Computed Tomography findings. Eur J Radiol 2016;85:1765–1772.

21.

Deng

Yang

Zhang

, et al. Predicting lymphovascular invasion in N0 stage non-small cell lung cancer: a nomogram based on Dual-energy CT imaging and clinical findings. Eur J Radiol 2024;179:111650.

22.

Tao

Shen

, et al. Solid-type adenocarcinoma on thin-section CT: quantitative parameters from dual-energy CT associated with spread through air spaces. Acta Radiol 2025;66:184–191.

23.

Wang

Yue

Fan

, et al. Spectral dual-layer computed tomography can predict the invasiveness of ground-glass nodules: a diagnostic model combined with thymidine kinase-1. J Clin Med 2023;12:1107.

24.

Chen

, et al. Dual-layer spectral detector CT: predicting the invasiveness of pure ground-glass adenocarcinoma. Clin Radiol 2022;77:e458–e465.

25.

Lin

Zhang

Suo

, et al. Correlation between dual-energy spectral CT imaging parameters and pathological grades of non-small cell lung cancer. Clin Radiol 2018;73:e411–412.e7.

26.

, et al. Histological subtypes of solid-dominant invasive lung adenocarcinoma: differentiation using dual-energy spectral CT. Clin Radiol 2021;76:77.e1–77.e7.

27.

Liu

Long

, et al. Predicting high-risk lung adenocarcinoma in solid and part-solid nodules on low-dose CT: a multicenter study. Acad Radiol 2025;32:2966–2976.

28.

Chen

Ding

Deng

, et al. Differentiating the invasiveness of lung adenocarcinoma manifesting as ground glass nodules: combination of dual-energy CT parameters and quantitative-semantic features. Acad Radiol 2024;31:2962–2972.

29.

Kim

Chung

, et al. Lung adenocarcinoma: CT features associated with spread through air spaces. Radiology 2018;289:831–840.

30.

Choe

Kim

Yun

, et al. Sublobar resection in stage IA non-small cell lung cancer: role of preoperative CT features in predicting pathologic lymphovascular invasion and postoperative recurrence. AJR Am J Roentgenol 2021;217:871–881.

31.

Ahn

Park

Jeon

, et al. Predictive CT features of visceral pleural invasion by T1-sized peripheral pulmonary adenocarcinomas manifesting as subsolid nodules. AJR Am J Roentgenol 2017;209:561–566.

32.

Gao

, et al. Evaluating the necessity of lymph node sampling in lung adenocarcinoma with ground glass opacities. Surgery 2024;176:927–933.

33.

Zeng

Zhou

, et al. Peritumoral radiomics increases the efficiency of classification of pure ground-glass lung nodules: a multicenter study. J Cardiothorac Surg 2024;19:05.

34.

Chen

, et al. An artificial-intelligence lung imaging analysis system (ALIAS) for population-based nodule computing in CT scans. Comput Med Imaging Graph 2021;89:101899.

35.

Bankier

MacMahon

Colby

, et al. Fleischner society: glossary of terms for thoracic imaging. Radiology 2024;310:e232558.

36.

Nicholson

Tsao

Beasley

, et al. The 2021 WHO classification of lung tumors: impact of advances since 2015. J Thorac Oncol 2022;17:362–387.

37.

DeLong

Clarke-Pearson

. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837–845.

38.

Song

Hou

Zhang

, et al. Comparison of outcomes following lobectomy, segmentectomy, and wedge resection based on pathological subtyping in patients with pN0 invasive lung adenocarcinoma ≤ 1 cm. Cancer Med 2022;11:4784–4795.

39.

Nitadori

Bograd

Kadota

, et al. Impact of micropapillary histologic subtype in selecting limited resection vs lobectomy for lung adenocarcinoma of 2cm or smaller. J Natl Cancer Inst 2013;105:1212–1220.

40.

Baig

Razi

Weber

, et al. Lobectomy is superior to segmentectomy for peripheral high grade non-small cell lung cancer ≤ 2 cm. J Thorac Dis 2020;12:5925–5933.

41.

Huang

Petersen

. Tumour spread through air spaces is a determiner for treatment of clinical stage I non-small cell lung cancer: thoracoscopic segmentectomy vs lobectomy. Lung Cancer 2025;201:108438.

42.

Ziora

Skiba

Kiczmer

, et al. Ten-year observational study of patients with lung adenocarcinoma: clinical outcomes, prognostic factors, and five-year survival rates. J Clin Med 2025;14:2552.

43.

Yeh

Nitadori

Kadota

, et al. Using frozen section to identify histological patterns in stage I lung adenocarcinoma of ≤ 3 cm: accuracy and interobserver agreement. Histopathology 2015;66:922–938.

44.

Zhou

Villalba

Sayo

TMS

, et al. Assessment of the feasibility of frozen sections for the detection of spread through air spaces (STAS) in pulmonary adenocarcinoma. Mod Pathol 2022;35:210–217.

45.

Domingo

. Molecular electron density theory: a modern view of reactivity in organic chemistry. Molecules 2016;21:1319.

46.

Tatsugami

Higaki

Kiguchi

, et al. Measurement of electron density and effective atomic number by dual-energy scan using a 320-detector computed tomography scanner with raw data-based analysis: a phantom study. J Comput Assist Tomogr 2014;38:824–827.

47.

Tsuchiya

Doai

Usuda

, et al. Non-small cell lung cancer: whole-lesion histogram analysis of the apparent diffusion coefficient for assessment of tumor grade, lymphovascular invasion and pleural invasion. PLoS One 2017;12:e0172433.

48.

Lee

Kim

, et al. Solitary pulmonary nodules: dynamic enhanced multi-detector row CT study and comparison with vascular endothelial growth factor and microvessel density. Radiology 2004;233:191–199.

49.

Kawai

Shibamoto

Hara

, et al. Can dual-energy CT evaluate contrast enhancement of ground-glass attenuation? Phantom and preliminary clinical studies. Acad Radiol 2011;18:682–689.

50.

Spira

Neumeister

Spira

, et al. Assessment of tumor vascularity in lung cancer using volume perfusion CT (VPCT) with histopathologic comparison: a further step toward an individualized tumor characterization. J Comput Assist Tomogr 2013;37:15–21.

51.

Kaichi

Tatsugami

Nakamura

, et al. Improved differentiation between high- and low-grade gliomas by combining dual-energy CT analysis and perfusion CT. Medicine (Baltimore) 2018;97:e11670.

52.

Yang

Sun

, et al. Invasive pulmonary adenocarcinomas versus preinvasive lesions appearing as pure ground-glass nodules: differentiation using enhanced dual-source dual-energy CT. AJR Am J Roentgenol 2019;213:W114–W122.

53.

Karcaaltincaba

Aktas

. Dual-energy CT revisited with multidetector CT: review of principles and clinical applications. Diagn Interv Radiol 2011;17:181–194.

54.

Kim

Park

, et al. Application of dual-energy spectral computed tomography to thoracic oncology imaging. Korean J Radiol 2020;21:838–850.

55.

Fan

Yue

, et al. Predicting the invasiveness of mixed ground-glass nodules based on spectral computed tomography-derived parameters and tumor abnormal protein levels: development and validation of a model. Acad Radiol 2025;32:2990–3005.

56.

Miao

Zhang

Zou

, et al. Correlation in histological subtypes with high resolution computed tomography signatures of early stage lung adenocarcinoma. Transl Lung Cancer Res 2017;6:14–22.

57.

Zwirewich

Vedal

Miller

, et al. Solitary pulmonary nodule: high-resolution CT and radiologic-pathologic correlation. Radiology 1991;179:469–476.

58.

Noguchi

Morikawa

Kawasaki

, et al. Small adenocarcinoma of the lung. Histologic characteristics and prognosis. Cancer 1995;75:2844–2852.

59.

Liu

Yang

, et al. Radiomic and quantitative-semantic models of low-dose computed tomography for predicting the poorly differentiated invasive non-mucinous pulmonary adenocarcinoma. Radiol Med 2023;128:191–202.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.15 MB

0.00 MB