Deep Learning Image Reconstruction Algorithm for Quantitative Assessment of Low-Dose Biphasic Chest CT in Chronic Obstructive Pulmonary Disease

Abstract

Objective

To evaluate the impact of deep learning image reconstruction (DLIR) on quantitatively assessing emphysema, air trapping and small airway dysfunction in chronic obstructive pulmonary disease (COPD) using low-dose inspiratory–expiratory chest CT.

Methods

Sixty-nine COPD patients underwent low-dose inspiratory-expiratory chest CT scans and pulmonary function tests (PFT) were prospectively enrolled. The CT images were reconstructed using 50% adaptive statistical iterative reconstruction (ASiR-V), DLIR-high (DLIR-H), medium (DLIR-M), and low (DLIR-L) strengths. The volumes and its percentages (relative to whole lung) characterizing emphysema, air trapping and small airway dysfunction were quantified on the inspiratory-expiratory CT scans.

Results

The total dose-length product was 128.99 ± 39.00 mGy·cm. For all patients, emphysema parameters were lowest for DLIR-H and highest for ASiR-V; small airway dysfunction parameters were highest with DLIR-H and lowest with ASiR-V; air trapping parameters were lowest with ASiR-V; highest with DLIR-M. Emphysema parameters demonstrated moderate negative correlations with FEV₁/FVC (r = –0.570 to –0.649, all p < 0.001). Air trapping and small airway dysfunction parameters showed weak negative correlations with MEF25%, MEF50%, and MEF75% (r = –0.320 to –0.381, all p < 0.001). When differentiating GOLD I–II from III–IV, all parameters showed AUC values ranging from 0.69 to 0.76, without statistically differences among reconstructions (DeLong’s test, p > 0.05), while the optimal thresholds varied across reconstructions.

Conclusion

In low-dose inspiratory–expiratory chest CT, DLIR may alter the lung function-related CT parameters compared to ASiR-V, but does not affect their correlations with PFTs or their efficacies in GOLD grading.

Introduction

Chronic obstructive pulmonary disease (COPD), defined by persistent airflow limitation, is one of the leading causes of death worldwide.¹ Its prevalence and impact are expected to increase, with annual deaths from COPD and related conditions projected to exceed 5.4 million by 2050.^2,3 The various clinical respiratory symptom profiles of COPD have been demonstrated to reflect its heterogeneity and predict future exacerbation risk, which accompanied with lung function decline, reduced physical activity, and increased mortality.^4–6 The disease primarily comprises emphysema and small airway disease, with variations in their combination and severity leading to differing levels of lung function impairment among patients. Assessing these components is essential for guiding treatment decisions, prognosis, and tracking disease progression in COPD patients. Quantifying lung volume using dual-phase (inspiratory and expiratory) chest CT facilitates the assessment of several critical aspects of chronic obstructive pulmonary disease (COPD), including emphysema areas, small airways dysfunction, and air-trapping regions.^7,8 Notably, the quantitative parameters obtained from these scans are significantly correlated with dyspnea severity as assessed by clinical pulmonary function tests (PFT) and the Global Initiative for Chronic Obstructive Lung Disease (GOLD) classification. Given chest CT’s exceptional capacity to visualize various features, such as emphysema, bronchi, airways, coronary artery calcifications, pulmonary vasodilation, and body composition, the GOLD 2023 report recommends its use for COPD patients with severe disease.⁹ This endorsement suggests that the adoption of dual-phase chest CT in clinical practice for COPD is expected to increase in the future. However, it is crucial to consider the ongoing concerns regarding radiation exposure.

The increase in image noise associated with radiation dose reduction presents a significant challenge in radiology.^8,10 Recent studies have demonstrated that deep learning image reconstruction (DLIR) algorithms are more effective than conventional filtered back projection (FBP) and iterative reconstruction methods in reducing lung tissue noise and enhancing the visualization of malignant findings in low-dose chest CT scans.⁹ However, findings indicate that higher noise reduction levels in CT scans may lead to an underestimation of emphysema volume compared to traditional FBP approaches.¹¹ Additionally, emphysema volumes measured using the same algorithm can vary significantly across different radiation doses.¹² Furthermore, while no studies to date have applied DLIR in the context of expiratory-phase chest CT, existing medical literature supports the inclusion of both inspiratory and expiratory-phase scans in COPD evaluation. Nonetheless, no research has relied on PFT results as the exclusive reference standard. As a result, the clinical value of DLIR in quantifying emphysema and airway retention volumes during low-dose dual-phase chest CT remains uncertain.

Based on dual-phase low-dose chest CT, this study aimed to: 1) Explore the impact of reconstruction algorithms on the quantification of emphysema and air trapping, as well as their correlations with PFT indices; and 2) Assess the clinical significance of parameters obtained using various reconstruction algorithms in the context of GOLD staging for COPD.

Materials and Methods

Patient Recruitment

This prospective study received approval from the institutional Ethics Committee and was conducted in accordance with the Declaration of Helsinki. Patients clinically diagnosed with COPD, who were admitted to our hospital between October 2023 and December 2024, were prospectively enrolled and underwent dual-phase low-dose chest CT. Informed consent was obtained from all participants.

Inclusion criteria: (1) According to the GOLD guidelines, the diagnosis of COPD is based on a patient’s history of exposure to risk factors, clinical symptoms such as cough, sputum production, and breathlessness, along with other clinical data. A definitive diagnosis is made through PFT that show a post-bronchodilator FEV1/FVC ratio of less than 0.7. Patients must also be excluded for conditions that could cause similar symptoms or persistent airflow limitation, such as asthma, bronchiectasis, lung cancer, tuberculosis, interstitial lung diseases, and cardiovascular diseases like coronary heart disease and heart valve diseases, which can mimic COPD. Exclusion criteria: (1) poor image quality due to respiratory motion artifacts; (2) incomplete analysis and processing parameters of the imaging software; (3) bronchiectasis, history of lung surgery; and (4) PFT with an interval ≤1 month from the CT examination. Finally, as shown in Figure 1, 90 patients were initially selected for this study, with 69 patients finally included.

Figure 1.

Diagram showing the patient inclusion process.

General data including gender, age, height, weight, smoking status, and body mass index (BMI), as well as clinical diagnosis results, were recorded. Computed tomography dose index (CTDI) and dose-length product (DLP) were documented from the data generated post-examination.

Pulmonary Function Test

A spirometer (MasterScreen, Vyaire Medical GmbH, China) was utilized to conduct pulmonary function tests. This testing involved the measurement of the forced expiratory volume in one second to forced vital capacity ratio (FEV1/FVC), as well as assessing maximal instantaneous forced expiratory flow rates at 25%, 50%, and 75% of vital capacity (MEF25%, MEF50%, and MEF75%). Additionally, the patient’s GOLD classification was determined based on the FEV1 metric as per the [Global Initiative for Chronic Obstructive Lung Disease 2023 Report].³ In this classification, GOLD I and II were combined into a single category designated as GOLD I–II, while GOLD III and IV were grouped together as GOLD III–IV.

CT Acquisition

A 256-row CT scanner (Revolution CT, GE HealthCare, USA) was utilized to scan patients from the thoracic inlet to the diaphragm. Prior to scanning, patients were instructed in breath administration and positioned supine. They were guided to elevate both upper limbs and hold their breath at the end of maximal inspiration and expiration during the respective inspiratory and expiratory phases of the scan. The scanning parameters included a tube voltage of 100 kV, an automatic tube current range of 50–350 mA, a preset noise index (NI) of 16, a pitch of 0.992, and a rotation speed of 0.35 s per rotation. Raw data were reconstructed using four different algorithms: adaptive statistical iterative reconstruction-Veo with a 50% weighting (ASiR-V) and deep learning image reconstruction with three levels (DLIR-H/M/L). The reconstruction employed a standard kernel with a layer thickness of 1.25 mm.

Image Analysis

The analysis was performed by a radiologist with 5 years of experience in chest imaging diagnosis using the Thoracic VCAR software on the Advanced workstation (version 4.7, GE HealthCare, USA). The low attenuation area (LAA) was defined as the volume below a specific HU threshold. The lung volume based on the CT attenuation value was divided into three categories: (1) lung attenuation area below −950 HU during the inspiratory phase was considered as emphysema volume (LAAisp-950), with LAAexp-950% representing the proportion percentage of emphysema from total lung volume; (2) lung attenuation area below −856 HU on expiratory CT was considered air-trapping volume (LAAexp-856), with LAAexp-856% representing the proportion percentage of air-trapping area from total lung volume; (3) lung attenuation area between −950 and −856 HU on expiratory CT considered small airway dysfunction volume (LAAexp-856~-950), with LAAexp-856~-950% representing the proportion percentage of small airway dysfunction from total lung volume.^8,12,13

Statistical Analysis

Statistical analysis was performed using SPSS 26.0 software. Quantitative data were expressed as mean ± standard deviation (SD), and categorical data were expressed as number (percentage). The comparison of parameters among the four reconstruction groups was conducted using the Friedman test. Correlation analysis between CT quantitative parameters and PFT indicators was performed using Spearman’s rank correlation, with the following grading: ≥0.8 very strong, 0.6–0.8 strong, 0.4–0.6 moderate, 0.2–0.4 weak, and <0.2 very weak. The diagnostic efficacy of each CT quantitative parameter in differentiating between GOLD grades I–II and III–IV was evaluated using the receiver operating characteristic (ROC) curve, and the area under the curve (AUC), optimal threshold, sensitivity, and specificity were calculated. The optimal threshold was determined based on the maxima Youden index, which was calculated as sensitivity+specificity-1.¹⁴ The difference in AUC between reconstruction algorithms was compared using the Delong’s test.^15,16 A two-tailed p< 0.05 was considered statistically significant. A “≈” represent the differences between the former and latter was not statistically significant (P>0.05), while a “<” or a “>” meant a statistically significant difference (P<0.05).

Results

Patient Characteristics and Radiation Dose

A total of 69 patients were included in this study, consisting of 9 females. The mean age was 62 ± 1.13 years (range: 36–79 years), with a mean BMI of 22.02 ± 0.47 (range: 19.80–25.55 kg/m²). Among them, 31 patients had smoking history. The distribution of GOLD classification was as follows: GOLD I (11 patients), GOLD II (16 patients), GOLD III (27 patients), and GOLD IV (15 patients). In the inspiratory phase, the mean CTDI was 1.58±0.40 mGy and, the mean DLP was 65.00±19.43 mGy·cm. In the expiratory phase, the mean CTDI was 1.58±0.39 mGy and the mean DLP was 64.00±19.65 mGy·cm (Table 1).

Table 1.

Patient Characteristics and Radiation Doses

Characteristics	Value
Age (year)	62.55±1.13
Height (m)	1.66±0.57
Weight (kg)	60.93±1.32
Body mass index (kg/m2)	22.02±0.47
Smoking (year)	20(7.5, 32.5)
Sex [Male/Female (%)]	52 (75.4%)/17(24.6%)
GOLD Grade	69
GOLD I	11(15.9%)
GOLD II	16(23.3%)
GOLD III	27(39.1%)
GOLD IV	15(21.7%)
Radiation dose
Inspiratory phase
CTDI (mGy)	1.58±0.40
DLP (mGy·cm)	65.00±19.43
Expiratory phase
CTDI (mGy)	1.55±0.39
DLP (mGy·cm)	64.00±19.65
PFT results
FEV1/FVC (%)	50.80(32.87,65.33)
FEV1-pre (L)	68.00(43.65,83.00)
MEF25% (L/s)	41.40(14.90,62.50)
MEF50% (L/s)	42.50(14.60,70.20)
MEF75% (L/s)	44.80(20.60,75.00)

Comparison of Quantitative Parameters Among the Four Reconstruction Algorithms Based on Low-Dose Inspiration–Expiration Chest CT

For a certain reconstruction algorithm, all parameters were significantly different between GOLD grades (P < 0.05). GOLD III–IV patients had values 23% to 41% higher for all parameters compared to those GOLD I–II patients (P < 0.05). Comparison of the reconstructed images revealed the following ranking: for LAAisp-950, ASIR-V50% > DLIR-L > DLIR-M ≈ DLIR-H; for LAAisp-950%, ASIR-V50% > DLIR-L > DLIR-M > DLIR-H. LAAexp-856 comparison: DLIR-M ≈ DLIR-H > DLIR-L > ASIR-V50%; LAAexp-856% comparison: DLIR-M > DLIR-L > DLIR-H ≈ ASIR-V50%; LAAisp-856~-950 and LAA-856~-950% comparisons: DLIR-H > DLIR-M > DLIR-L > ASIR-V50. As shown in Figure 2.

Figure 2.

Histogram of quantitative CT parameters of GOLD grouping in different reconstruction modes. (A) Volume of LAAisp-950; (B) Volume percentage of LAAisp-950; (C) Volume of LAAexp-856; (D) Volume percentage of LAAexp-856; (E) Volume of LAAexp-856~-950; (F) Volume percentage of LAAexp-856~-950. * p < 0.05, ** p < 0.01, *** p < 0.001; I, GOLD Stage I (mild airflow limitation); II, GOLD Stage II (moderate airflow limitation); III, GOLD Stage III (severe airflow limitation); IV, GOLD Stage IV (very severe airflow limitation).

Correlations Between CT Parameters and PFT Indexes

As shown in Table 2, all CT parameters had significant and negative correlations with corresponding PFT results, across the four reconstruction algorithms (all P<0.05). In detail, for inspiratory phase, DLIR-L and DLIR-M parameters had the strongest correlations with FEV1/FVC at LAAisp-950 (ρ=−0.607) and LAAisp-950% (ρ=−0.649), respectively, while ASiR-V50% had the weakest correlations (ρ=−0.549 for LAAisp-950, ρ=−0.587 for LAAisp-950%). For expiratory phase, all CT parameters had weak correlations with MEF-related PFT results (ρ=−0.320~-0.392). Different algorithms lead to changes in correlation coefficient less or equal than 0.02 for following pairs: LAA-exp856 vs MEF25% (0.009), LAA-exp856 vs MEF50% (0.016), LAA-exp856 vs MEF75% (0.02), LAA-exp856% vs MEF25% (0.011), LAA-exp856-950 vs MEF25% (0.009), LAA-exp856-950 vs MEF50% (0.016), LAA-exp856-950 vs MEF75% (0.02), LAAexp856-950% vs MEF25% (0.011). While for LAA-exp856% vs MEF50%, LAA-exp856% vs MEF75%, LAAexp856-950% vs MEF50% and LAAexp856-950% vs MEF75%, DLIR-M had the lowest correlations, while ASiR-V50% showed the strongest, with the correlation coefficient changes of 0.037, 0.044, 0.037, 0.044, respectively.

Table 2.

Correlation Analysis Between Pulmonary Function Test and CT Lung Quantification

Reconstruction	Quantitative Parameters vs PFT
	LAAisp-950 vs FEV1/FVC				LAAisp-950 vs FEV1/FVC
	ρ	P			ρ	P
DLIR-H	−0.584	<0.001			−0.616	<0.001
DLIR-M	−0.57	<0.001			−0.649	<0.001
DLIR-L	−0.607	<0.001			−0.626	<0.001
ASIR-V50%	−0.549	<0.001			−0.587	<0.001
	LAA-exp856 vs MEF25%		LAA-exp856 vs MEF50%		LAA-exp856 vs MEF75%
	ρ	P	ρ	P	ρ	P
DLIR-H	−0.365	0.002	−0.344	0.004	−0.341	0.004
DLIR-M	−0.365	0.002	−0.329	0.006	−0.323	0.004
DLIR-L	−0.367	0.002	−0.341	0.004	−0.343	0.004
ASIR-V50%	−0.358	0.003	−0.345	0.004	−0.343	0.004
	LAA-exp856% vs MEF25%		LAA-exp856% vs MEF50%		LAA-exp856% vs MEF75%
	ρ	P	ρ	P	ρ	P
DLIR-H	−0.379	0.001	−0.367	0.002	−0.328	0.006
DLIR-M	−0.370	0.002	−0.355	0.003	−0.320	0.008
DLIR-L	−0.381	0.001	−0.379	0.001	−0.340	0.002
ASIR-V50%	−0.376	0.001	−0.392	0.001	−0.364	0.002
	LAA-exp856-950 vs MEF25%		LAA-exp856-950 vs MEF50%		LAA-exp856-950vs MEF75%
	ρ	P	ρ	P	ρ	P
DLIR-H	−0.365	0.002	−0.344	0.004	−0.341	0.004
DLIR-M	−0.365	0.002	−0.329	0.006	−0.323	0.007
DLIR-L	−0.367	0.002	−0.341	0.004	−0.343	0.004
ASIR-V50%	−0.358	0.003	−0.345	0.004	−0.343	0.004
	LAAexp856-950% vs MEF25%		LAAexp856-950% vs MEF50%		LAAexp856-950% vs MEF75%
	ρ	P	ρ	P	ρ	P
DLIR-H	−0.379	0.001	−0.367	0.002	−0.348	0.006
DLIR-M	−0.370	0.002	−0.355	0.003	−0.320	0.008
DLIR-L	−0.381	0.001	−0.379	0.001	-0.364	0.002
ASIR-V50%	−0.376	0.001	−0.392	0.001	−0.364	0.002

Diagnostic Efficacy of CT Quantitative Parameters in GOLD Classification

As shown in Table 3, all CT quantitative parameters were statistically significant in distinguishing GOLD I–II and III–IV under the four reconstruction algorithms (AUC ranging from 0.69–0.76, all P < 0.05). Delong’s test showed that the differences of AUCs among different reconstruction algorithms were not statistically significant (P > 0.05).

Table 3.

ROC Analysis of CT Quantitative Parameters in GOLD Classifications

Reconstruction		CT Quantitative vs GOLD Classifications
	LAAisp-950
	AUC (95% CI)	P	Cutoff	Yoden index	Sensitivity	Specificity
DLIR-H	0.70(0.57,0.82)	0.005	0.137	0.421	69.4%	72.7%
DLIR-M	0.70(0.57,0.82)	0.005	0.081	0.429	61.1%	81.8%
DLIR-L	0.69(0.57,0.82)	0.006	0.196	0.425	66.7%	75.8%
ASIR-V50%	0.71(0.59,0.84)	0.002	0.222	0.369	61.1%	75.8%
	LAAisp-950%
DLIR-H	0.70(0.57,0.83)	0.004	1.867	0.457	63.9%	81.8%
DLIR-M	0.70(0.57,0.83)	0.004	2.043	0.485	66.7%	81.8%
DLIR-L	0.69(0.57,0.82)	0.006	2.26	0.374	55.6%	81.8%
ASIR-V50%	0.70(0.58,0.82)	0.004	4.807	0.394	66.7%	72.7%
	LAAexp-856
DLIR-H	0.72(0.60,0.85)	0.001	2.261	0.515	66.7%	84.8%
DLIR-M	0.73(0.61,0.85)	0.001	2.242	0.455	66.7%	78.8%
DLIR-L	0.74(0.62,0.86)	0.001	2.171	0.455	66.7%	78.8%
ASIR-V50%	0.73(0.61,0.85)	0.001	2.043	0.54	72.2%	81.8%
	LAAexp-856%
DLIR-H	0.71(0.58,0.83)	0.003	48.474	0.382	80.6%	57.6%
DLIR-M	0.73(0.61,0.85)	0.001	46.971	0.407	52.8%	87.9%
DLIR-L	0.75(0.63,0.86)	<0.001	53.411	0.475	77.8%	69.7%
ASIR-V50%	0.76(0.64,0.87)	<0.001	50.919	0.505	77.8%	72.7%
	LAAexp-856~-950
DLIR-H	0.72(0.59,0.84)	0.002	2.130	0.477	62.9%	84.8%
DLIR-M	0.73(0.60,0.85)	0.001	2.105	0.477	62.9%	84.8%
DLIR-L	0.72(0.60,0.85)	0.001	2.106	0.505	65.7%	84.8%
ASIR-V50%	0.72(0.60,0.85)	0.001	2.043	0.532	71.4%	81.8%
	LAAexp-856~-950%
DLIR-H	0.73(0.61,0.85)	0.001	35.025	0.339	40.0%	93.9%
DLIR-M	0.73(0.60,0.85)	0.001	41.885	0.393	51.4%	87.9%
DLIR-L	0.72(0.60,0.84)	0.002	41.998	0.393	51.4%	87.8%
ASIR-V50%	0.72(0.60,0.840)	0.002	41.118	0.419	57.1%	84.8%

However, the optimal discrimination threshold for a particular measurement varied among reconstruction algorithms: LAAisp-950 and LAAisp-950% had AUCs ranging from 0.69–0.71 and 0.69–0.70 among reconstruction algorithms, respectively. The optimal thresholds for DLIR images were lower than those of ASiR-V50%, and DLIR-H < DLIR-M < DLIR-L for LAAisp-950%, DLIR-M < DLIR-H < DLIR-L for LAAisp-950.

The AUC values were 0.72–0.74 and 0.71–0.76 for LAAexp-856 and LAAexp-856%, respectively, and the optimal thresholds for LAAexp-856: ASiR-V50% < DLIR-L < DLIR-M < DLIR-H, whereas for LAAexp-856%, the optimal DLIR thresholds are lower than ASiR-V50% and the DLIR-M threshold is the lowest.

The AUC values were both 0.72–0.73 for LAAexp-856~-950 and LAAexp-856~-950%, respectively, and the optimal thresholds for LAAexp-856~-950: DLIR-H < DLIR-L < DLIR-M < ASiR-V50%, whereas the optimal thresholds for LAAexp-856~-950%: DLIR-H< ASiR-V50% < DLIR-M < DLIR-L.

Discussion

This study evaluated the utility of DLIR algorithm in low-dose dual-phase chest CT for the quantitative assessment of emphysema, air trapping, and small airway dysfunction in patients with COPD. The findings indicate that, compared to the ASIR-V 50% algorithm, DLIR tends to underestimate emphysema measurements while overestimating air trapping and small airway dysfunction parameters. DLIR may enhance the correlation between quantitative emphysema indices and the FEV1/FVC ratio; however, it does not significantly affect the correlations of air trapping or small airway dysfunction parameters with MEF25%, MEF50%, or MEF75%. Furthermore, although DLIR does not alter the overall performance of quantitative parameters in distinguishing between GOLD grades I–II and III–IV, it may shift the optimal threshold for this differentiation.

Reconstruction algorithms had impacts on the quantitative parameters derived from low-dose dual-phase chest CT. In this study, DLIR-H produced the lowest values for LAAisp-950 and LAAisp-950%, while ASIR-V50% resulted in the highest values for these parameters. This is similar to the findings of Ferri.¹⁷ This discrepancy can be attributed to the fact that the quantification of emphysema relies on the CT density histogram, whose shape is significantly influenced by the image noise level and the filtering effects of the reconstruction algorithm.¹⁸ Wisselink^10,19 confirmed that the radiation dose level and noise reduction methods (including DLIR) affect the characteristics of CT density histograms. They further noted that at ultra-low doses, both iterative reconstruction and DLIR algorithms led to bias and variability in emphysema measurements. Their results and our studies all emphasize the need for standardizing scanning and reconstruction parameters, as well as accounting for algorithms, to obtain consistent quantification. Notably, DLIR in this study overestimating the values of LAAexp-856, LAAexp-856%, LAAexp-856~-950, and LAAexp-856~-950% compared to ASIR-V. This also highlights the different performances of reconstruction algorithms in quantifying airway trapping and small airway dysfunction, which has not been addressed in prior studies.

The quantitative parameters derived from low-dose inspiratory-expiratory chest CT showed significant correlations with the PFT indices across different reconstruction algorithms, although the strength of these correlations varied depending on the algorithm. LAAisp-950 and LAAisp-950%, reflecting the degree of airflow limitation, exhibited a moderate-to-strong negative correlation (ρ = −0.549~-0.649) with FEV₁/FVC, in line with previous studies^19,20 Notably, while previous study used conventional-dose chest CT, the present study employed low-dose chest CT, demonstrating the feasibility of quantifying emphysema volume with low-dose CT.¹⁰ More importantly, we found that LAAisp-950 (DLIR-L) and LAAisp-950% (DLIR-M) measured from DLIR images were strongly correlated with FEV1/FVC (ρ = −0.607 and ~−0.647). Previous studies have shown that the agreement between emphysema measurements from low-dose DLIR images and those from standard-dose FBP images is higher than the agreement between FBP and ASIR-V.^21,22 These suggest that, while FBP is often considered the “gold standard”²³ for reflecting true anatomical structure, under low-dose conditions, DLIR algorithms (especially those at low- and medium- strengths) may produce quantitative data that better reflect true anatomy or correlate more closely with functional indices. Moreover, we chose clinical gold standard PFT as the reference for correlation analysis, making the results more clinically relevant.

Air trapping (LAAexp-856, LAAexp-856%) and small airways dysfunction (LAAexp-856~-950, LAAexp-856~-950%) parameters showed weak negative correlations (ρ = −0.320 ~ −0.392) with small airway function indicators on the PFT (MEF25%, MEF50%, MEF75%).²⁴ The overall correlation remains weak, which may be due to that CT measures regional air-trapped volumes, while MEF reflects overall small airway function. Additionally, the enhancement of marginal structures may influence the quantitative accuracy. Notably, the small differences between LAAexp-856 and LAAexp-856~-950, and between LAAexp-856% and LAAexp-856~-950%, along with their similar correlations with MEF, suggest that in the expiratory phase, abnormal air trapping volume in COPD is primarily distributed in the density range of −856 HU to −950 HU.

The small airways dysfunction metrics we measured (LAAexp-856~-950 and LAAexp-856~-950%) were not based on registered spiration-respiration images, but only from the expiratory phased images. Surprisingly, our results showed that the unmatched parameters exhibited a similar or even superior trend in correlations between CT parameters and PFT results, when compared to Hwang HJ’s¹⁰ registered results. This may be explained by the different populations the two studies enrolled, as we included confirmed COPD patients, of whom small airway trapping is common. Nevertheless, our findings indicated that, in the absence of airway-registered calibration, in COPD patients, measuring the above small airway retention-related indices in the expiratory phase may be effective.

All CT quantitative parameters demonstrated good performances in distinguishing between GOLD grades I–II and III–IV patients, although no statistically significant differences in AUC were observed between the algorithms, the optimal thresholds varied by algorithm. This underscores the need to apply specific thresholds for each reconstruction algorithm in practice. The diagnostic performance differed noticeably before and after algorithm-specific threshold correction. When the LAAisp-950 fixed cut-off value of ASIR-V50% was applied to DLIR series without adjustment, DLIR-H had a sensitivity of 63.9% and a specificity of 72.7%, DLIR-M presented a sensitivity of 55.6% and a specificity of 81.8%, and DLIR-L showed a sensitivity of 69.4% and a specificity of 66.7%. After adopting the optimal cut-off values determined by the maximum Youden’s index for each DLIR algorithm, the overall diagnostic balance was significantly optimized. The sensitivity of DLIR-H increased to 69.4% while its specificity remained stable at 72.7%. DLIR-M gained a higher sensitivity of 61.1% with its specificity kept at 81.8%. For DLIR-L, the specificity was markedly elevated to 75.8%, with only a slight drop in sensitivity to 66.7%. Therefore, simply using a universal threshold across different reconstruction algorithms leads to suboptimal diagnostic performance. Algorithm-specific threshold correction effectively compensates for the numerical bias of CT quantitative parameters introduced by DLIR reconstruction. It may achieve a better balance between sensitivity and specificity, and maintain consistent and reliable diagnostic efficacy for distinguishing mild-to-moderate (GOLD I–II) and severe-to-very severe (GOLD III–IV) COPD in routine clinical practice.

Furthermore, as GOLD I–II and III–IV exhibit different clinical progressions, low-dose inspiration-expiration chest CT may be an effective tool to assist disease monitoring. Additionally, the total radiation dose of dual-phase chest CT in this study was kept low (mean total DLP = 136.44 mGy·cm), much lower than most national diagnostic reference levels for conventional chest CT scans. Therefore, this low-dose protocol is especially suitable for (1) patients unable to complete standardized PFTs due to acute exacerbations or other factors, and (2) patients requiring long-term follow-up and multiple assessments of lung structural changes and disease progression.

The present study has the following limitations: (1) Its single-center design and relatively small sample size may compromise the stability of the results; (2) The optimal Hounsfield unit (HU) thresholds for quantifying emphysema and air trapping under the low-dose conditions employed in this study were not investigated. Notably, previous studies have indicated that low-dose thresholds may differ from those used under conventional doses; (3) The absence of a conventional-dose CT control group precluded the direct quantification of quantitative bias induced by low dose and different reconstruction algorithms; (4) The impact of image quality—including noise, texture, and artifacts—of different reconstruction algorithms on quantification accuracy and diagnostic confidence has not been systematically evaluated; (5) The correlation between CT air-trapping parameters and small airway function indicators requires further investigation. Future studies should integrate more detailed CT parameters related to airway lumen and wall; (6) Although obese patients were not intentionally excluded, the included participants had a narrow BMI range (19.80–25.55 kg/m²), limiting the generalizability of the results to obese populations; (7) Although volumes related to emphysema and air trapping were calculated separately for the inspiratory and expiratory phases, inspiratory-expiratory volume changes were not quantified via airway registered, and parameter-response mapping was not conducted; (8) Current results are unable to make a definitive recommendation for a preferred DLIR strength level for routine clinical use. A comprehensive study integrating image quality, subjective evaluation and quantitative analysis should be proceeded further.

In conclusion, in low-dose (total DLP <140 mGy·cm) dual-phase chest CT scans, DLIR underestimates emphysema parameters while overestimating air-trapping and small airway dysfunction parameters, and may maintain or improve the correlations between CT parameters and PFT results. With no affected efficacy in differentiating GOLD grades, DLIR images may alter the optimal threshold for differentiation. An expanded sample size is needed for further clinical validation in the future.

Funding Statement

This study was sponsored by the Fujian Provincial Health Technology Project (Grant number: 2024TG004), The Joint Funds for the Innovation of Science and Technology, Fujian province (Grant number: 2024Y9322)and Startup Fund for scientific research, Fujian Medical University (Grant number: 2023QH1048).

Abbreviations

CT, Computed Tomography; COPD, chronic obstructive pulmonary disease; ASiR-V, adaptive statistical iterative reconstruction; DLIR, deep learning image reconstruction; PFT, pulmonary function test; GOLD, global initiative for obstructive lung disease; HU, Hounsfield unit; LAAisp-950, lung attenuation area below −950 HU during the inspiratory phase; LAAisp-950%, percentage of lung attenuation Area below −950 HU during the inspiratory phase; LAAexp-856, lung attenuation area below-856 HU on expiratory CT; LAAexp-856%, percentage of lung attenuation area below −856 HU on expiratory; LAAexp-950~-856, lung attenuation area between-950 to −856 HU on expiratory CT; LAAexp-950~-856%, percentage of lung attenuation area between-950 to −856 HU on expiratory CT; FEV1/FVC, forced expiratory volume in one second to forced vital capacity; MEF, maximal instantaneous forced expiratory flow rates; CTDI, computed tomography dose index; DLP, dose-Length product.

Data Sharing Statement

All data generated or analyzed during this study are included in this published article. As further research proceeds, raw and processed data are not available to all readers. Moreover, these data would be provided by Y. Liu upon reasonable request.

Ethics Approval

The study was conducted in accordance with the Declaration of Helsinki. It was also approved by the Ethics Committee of Union Hospital Affiliated to Fujian Medical University (No. 2023QH016), and the ethical document has been issued.

Consent for Publication

All authors confirm that they have obtained the necessary written consent for publication from all study participants. The participants have been shown the contents of the article to be published and have agreed to the publication of all relevant images and data included in this study. The authors will provide copies of the signed consent forms to the editorial office upon request.

Author Contributions

Liwei Xue, Qiong Lin and Xiongxin Ye contributed equally to this work and share the first authorship. Yunjing Xue and Yuanfen Liu contributed equally as corresponding authors. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agreed to be accountable for all aspects of the work.

Disclosure

All authors declare no competing interests.