Segmentation-based deep learning emphysema quantification using chest CT: improved accuracy and robustness vs LAA-950

Abstract

Objectives

To develop a deep learning segmentation algorithm to enable accurate, reliable emphysema quantification while improving agreement with radiologist assessments and pulmonary function tests.

Materials and methods

The model was developed using retrospective virtual and clinical datasets. Virtual data enabled pre-training using ground truth across controlled parameters, including scanners, doses, and reconstruction kernels, while clinical data enabled fine-tuning with expert-annotated emphysema masks. Segmentation accuracy was quantified using the Dice coefficient, and robustness was quantified using emphysema percentage consistency across imaging conditions. Model-based emphysema percentage was correlated with Fleischner visual scores (ordinal; 0–5) and pulmonary function tests (DLCO, FEV₁pp, and FEV1/FVC) and compared to LAA-950. Statistical analysis included univariate/multivariate correlations. Quantitative assessment included Dice, bias, limits of agreement, and reproducibility coefficient.

Results

Virtual data included 20 human models (mean age: 43 years ± 11 [SD], 10 men), and clinical data included multi-center cohorts of 101 patients (C1; 57 years ± 8, 54 men), 23 patients (C2; 57 years ± 7, 14 men), and 1159 patients (C3; 65 years ± 9, 586 men). The model outperformed LAA-950 in segmentation accuracy, achieving higher Dice scores across virtual (76.6% ± 8.8 vs 51.5% ± 23.5), C1 (48.4% ± 24.2 vs 22.5% ± 20.2), and C2 (64.8% ± 9.3 vs 32.7% ± 16.8) cohorts. Furthermore, analysis showed improved bias and limits of agreement (C1; 1.0% ± 4.3 vs 2.1% ± 15.1) and stronger correlations with visual scoring (C1; 0.77 vs 0.47) and pulmonary function tests (C3; DLCO multivariate R²: 0.31–0.32 vs 0.21–0.25).

Conclusion

The proposed model provided a promising, clinically relevant alternative to LAA-950, improving accuracy and consistency in emphysema quantification and further aligning with radiologist assessments and lung function metrics.

Graphical Abstract

Introduction

Pulmonary emphysema is a chronic lung disease characterized by the irreversible destruction of alveolar walls and abnormal enlargement of distal airspaces, ultimately leading to impaired respiratory function. One way to evaluate these structural alterations in vivo is through computed tomography (CT), which allows for detailed visualization of lung parenchyma. This visualization ability has led to the development of quantitative CT (QCT) techniques, which can objectively measure the extent and progression of emphysema. One such QCT metric is LAA-950, defined as the percentage of lung voxels with attenuation below −950 Hounsfield Units (HU) on inspiratory CT scans. LAA-950, together with other density-based biomarkers such as Perc15 (15th percentile of lung density histogram), has been increasingly used in clinical trials as an endpoint to evaluate disease progression [1]. This growing use is driven by their ability to detect subtle structural changes in the lung—even when lung function measurements like spirometry remain within normal ranges [2]—and their greater sensitivity compared to conventional physiological and health status indices [3]. However, despite their promise, these biomarkers are vulnerable to variability introduced by scanner models [4], imaging settings (e.g., radiation dose, reconstruction kernel, slice thickness) [5], and patient-related factors, including breath-hold level [2, 6]. Such variability can undermine the reliability of emphysema quantification, ultimately creating a discordance between imaging biomarkers and the actual disease condition.

Previous studies have proposed different solutions to mitigate variability in emphysema quantification, including image harmonization using analytical or deep learning techniques [7–11], linear mixed-effect models [12], analytical adjustments of lung volume, noise, and bias [13], and segmentation methods such as the hidden Markov measure field (HMMF) model [14] and 2D deep learningbased approaches [15]. However, some methods provide single attribute corrections (e.g., harmonization for reconstruction kernel effect), thus being limited to the parameter they cover. Others fail to generate a valid emphysema mask for the potential subtyping task, need manual tuning of their parametric models for newer scans [14], or lack ground-truth emphysema masks to be used as a reliable reference [15].

The purpose of this study was to develop a segmentation-based deep learning algorithm for automated and reliable emphysema quantification across scanner models, settings, and repeated acquisitions.

Materials and methods

Figure 1 outlines the retrospective study design incorporating both virtual and clinical datasets.

Fig. 1.

An overview of the study design

Study datasets

Virtual dataset

This study used data from 20 virtual patient models generated using a previously established framework [16–21], combining anthropomorphic phantoms based on real human CT images—capturing variation in sex, size, body habitus, and lung volume—with anatomically informed models of lung structures, including airways, vasculature, and secondary pulmonary lobules [19, 20]. Emphysema models were derived from COPDGene CT data, representing a range of subtypes (centrilobular and paraseptal) and variability in distribution, size, and density (a summary of patient attributes provided in Table 1 of [21]), and were manually segmented and incorporated into the human models [21]. The human models were imaged using a scanner-specific CT simulator (DukeSim) [22, 23], configured to emulate one photon-counting CT (PCCT: NAEOTOM Alpha) and two energy-integrating CT (EICT: SOMATOM Definition Flash and SOMATOM Force) scanners (Siemens Healthineers). DukeSim takes a virtual patient and acquisition protocol as input and generates realistic projection data, and has been validated for accuracy against experimental measurements across scanner technologies and imaging parameters [23–25].

Table 1.

Participant demographics of virtual data

Parameter	Value
No. of participants	20
Demographic information
Age (years)	43 ± 11
Age range (years)	27–59
Sex	10 male (50%), 10 female (50%)
BMI (kg/m2)	25.2 ± 6.0
Imaging biomarkers
Total lung capacity (L)	5.3 ± 1.2

For this study, DukeSim was configured to simulate CT scans at three radiation dose levels (CTDI_vol: 1.3, 6.5, 13 mGy) and three reconstruction kernels (soft, medium, sharp) per scanner. The projection data were reconstructed using the vendor-specific reconstruction software (ReconCT, Siemens Healthineers) with 1.50 mm slice thickness and 0.98 mm × 0.98 mm pixel size (512 × 512 matrix, 500 mm reconstruction field of view). This resulted in a virtual dataset of 540 CT volumes [5].

Clinical dataset

This study included data from the Genetic Epidemiology of COPD (COPDGene) study, a large-scale multi-center observational cohort conducted at 21 centers across the United States (www.clinicaltrials.gov [NCT00608764]) [26], approved by the institutional review board and compliant with the Health Insurance Portability and Accountability Act, with written informed consent provided from all participants. Clinical data were de-identified, and CT data were included from phases 1 and 2 of the COPDGene study in different cohorts (named C1–C3). Phase 1 data were acquired at full-dose (200 mAs) and used for model fine-tuning and validation (C1 cohort; 101 subjects reconstructed with soft and sharp kernels), and independent testing (C2 cohort;23 subjects reconstructed only with a soft kernel). Phase 2 data included a subset of subjects with full-dose and reduced-dose acquisitions (40–80 mAs with dose level varying due to tube current modulation) [27] and were used to evaluate the variability of emphysema quantification across imaging conditions (C3 cohort; 1159 subjects reconstructed with soft and/or sharp kernels). All data included lung segmentation using Thirona lung quantification software (LungQ) [28] and were reconstructed using a weighted filtered back projection algorithm. Table S1 summarizes the clinical scanner make and models used in this study.

Training and evaluation methodology

Emphysema is a diffuse pulmonary disease, making manual annotations highly time-consuming for expert radiologists, taking up to 40 h per single case [29]. To address this challenge, we employed a training strategy leveraging virtual imaging datasets, which provide ideal reference emphysema masks unobtainable in clinical data due to the absence of true ground truth.

A 3D UNet-based emphysema segmentation model (herein referred to as the AI-based model) was first trained on virtual data, allowing systematic evaluation of architectures and fine-tuning of hyperparameters. The dataset was split on a patient-level basis into training (60%), validation (20%), and testing (20%) subsets.

Despite the advantages of virtual imaging data (e.g., annotated ground truth, controlled acquisition parameters), models trained solely on such data may underperform on real clinical scans due to subtle differences in noise, texture, and anatomy. To address this, we applied a human-in-the-loop continual learning strategy. The best-performing model from the initial stage generated masks on clinical scans. These masks were then refined by a reader (radiology resident with 1 year of experience) using 3D Slicer and fed back into the model for fine-tuning. During this process, the reader consulted with a more experienced radiologist (with > 5 years of post-training experience) in multiple sessions to help ensure consistency and accuracy. This iterative process was performed over two data batches from the COPDGene phase 1 dataset (101 total cases; N = 51 and 50), improving model accuracy and robustness on clinical data. This approach was motivated by the heterogeneous morphology of emphysema on CT, which makes manual, slice-by-slice annotation labor-intensive. To further reduce annotation burden, CT data were downsampled two-fold along the z-axis, halving the number of slices. To preserve segmentation continuity, the corresponding masks were subsequently interpolated along the z-axis to resample them to the original slice spacing by estimating intermediate label values between adjacent slices. This step reduced discontinuities and mitigated jagged boundaries in the segmentation masks. The resampled masks were then used consistently for model training, inference, and performance evaluation.

Five-fold cross-validation was performed on all cases in the C1 cohort to assess model improvement and adaptation to clinical data, particularly given that initial training was conducted on virtual datasets. Acknowledging the need for independent validation, each of the five trained models was subsequently tested on cases from the C2 cohort, with ground truth emphysema annotations performed from scratch on selected CT slices. The slices were uniformly sampled across the 10th–90th percentile of the cranio-caudal lung extent to ensure representative coverage while avoiding regions with minimal lung area. The best-performing model was further evaluated for robustness in emphysema quantification across varying dose and reconstruction settings of the C3 cohort. All results from this section were compared against LAA-950, as well as LAA-910 measurements.

Lastly, segmentation was also applied to harmonized CT scans from the C3 cohort to assess the impact of image harmonization on quantification accuracy. Harmonization was performed using a physics-based method that standardizes images to a reference image quality index (i.e., noise and spatial resolution) [30]. This process aligns scans to a reference imaging condition of full-dose, soft-kernel reconstruction, given the superior accuracy of lung density biomarkers under soft-kernel settings [5, 31].

Correlation with CT visual scores and pulmonary function tests

The AI-based emphysema quantification was compared to LAA-950 based on alignment with expert interpretation and pulmonary function tests. For expert alignment, both methods were assessed against CT visual scores for centrilobular emphysema (0–5 scale, 0 = absent, 1 = trace, 2 = mild, 3 = moderate, 4 = confluent, 5 = advanced destructive) following the Fleischner Society classification [32, 33]. The objective was to assess the correlation of each quantification method with visual scores.

The clinical relevance of each emphysema quantification method—individually and in combination with other imaging biomarkers—was evaluated by assessing their association with pulmonary function tests. These tests included DLCO (diffusing capacity of the lungs for carbon monoxide, adjusted for hemoglobin and altitude), FEV1pp (percent predicted forced expiratory volume in one second), and FEV1/FVC (the ratio of forced expiratory volume in one second to forced vital capacity). Additional imaging biomarkers related to small airway disease and airway remodeling included PRM-fSAD (functional small airway disease based on parametric response mapping), Pi10 (standardized airway wall thickness for an airway with an internal perimeter of 10 mm), AWT (average airway wall thickness), and WallAreaPct (percentage of airway cross-sectional area occupied by the airway wall). These variables were quantified using Thirona lung quantification software (LungQ) and selected based on their established relevance in previous studies [34, 35].

Model architecture and training details

Figure S1 illustrates the proposed model based on a standard 3D-UNet architecture [36]. Training was performed using 3D patches of size 128 × 128 × 32 extracted from lung parenchyma regions, with intensity values normalized to [0, 1] using a window of [−1024, 1500] HU. The model was optimized to minimize Dice loss using the Adam optimizer and was trained for up to 500 epochs with a starting learning rate of 1e-4. A batch size of 16 (4 × NVIDIA RTX A5000 GPUs, 24 GB each) was used during training.

Statistical analysis

Model performance was evaluated using three metrics: 3D Dice coefficient, patient-level emphysema quantification variability, and reproducibility coefficient. The Dice coefficient quantified agreement between predicted and reference masks (ground truth for virtual data; expert-delineated for clinical data). Emphysema variability was defined as the difference in emphysema percentage between each imaging condition and the reference measurement (ground truth for virtual data; expert-delineated for C1 and C2 cohorts; full-dose, soft-kernel reconstructions for C3 cohort [5, 31]), assessed via Bland–Altman plots and their parameters, namely bias and limits of agreement (±1.96 SD). The reproducibility coefficient–representing the minimum detectable lung density change at 95% confidence–was estimated via nonparametric bootstrapping with 5000 resamples. Results were reported as mean ± standard deviation.

Alignment between quantitative metrics (LAA-950 and AI-based) and CT visual scores was evaluated using Spearman correlation, Kruskal–Wallis H-tests, and ordinal logistic regression. Model fit was compared using loglikelihood, Akaike (AIC), and Bayesian (BIC) information criterion, with higher log-likelihood and lower AIC/BIC indicating better fit. Correlation between emphysema metrics and pulmonary function tests was assessed using univariate and multivariate analyses. Univariate analysis was performed for each emphysema biomarker and imaging condition (three in total; full-dose with soft and sharp kernels, reduced-dose with soft kernel). Multivariate analysis included emphysema metrics and additional imaging biomarkers (PRM-fSAD, Pi10, AWT, and WallAreaPct). Predictors with variance inflation factor (VIF) > 10 or high Pearson correlations were excluded to reduce multicollinearity. All predictors were z-score normalized, and regression was performed using ordinary least squares (Python, statsmodels), with fit assessed by R², log-likelihood, and AIC. S.S.-P. performed all statistical analyses and served as the statistical guarantor of the study.

Results

Cohort characteristics

Tables 1 and 2 present the demographics of virtual and clinical datasets. In this study, the virtual population included 20 subjects (mean age: 43 years ± 11 years [standard deviation], 10 male). The clinical data included three cohorts (abbreviated as C1–C3): 101 participants from phase 1 for model fine-tuning (C1; mean age: 57 years ± 8 years [SD], 54 male), 23 participants from phase 1 for independent testing (C2; mean age: 57 years ± 7 years [SD], 14 male), and 1159 participants from phase 2 for variability analysis (C3; mean age: 65 years ± 9 years [SD], 586 male).

Table 2.

Participant demographics of COPDGene patients

Parameter	C1 cohort	C2 cohort	C3 cohort
	Independent testing		Variability analysis
	FD (n = 101)	FD (n = 23)	FD (n = 1159)	RD (n = 1159)
Age (years)	57 ± 8	57 ± 7	65 ± 9
Age range (years)	45–73	47–73	47–87
Sex
BMI	28.4 ± 6.0	28.0 ± 6.6	29.3 ± 6.5
Male	54 (53.5)	14 (60.9)	586 (50.6)
Female	47 (46.5)	9 (39.1)	573 (49.4)
Race
African American	29 (71.3)	10 (43.5)	420 (36.2)
White	72 (28.7)	13 (56.5)	739 (63.8)
Visit GOLD stage
PRISm	6	2	178
GOLD 0	28	5	528
GOLD 1	12	3	111
GOLD 2	26	8	193
GOLD 3	23	3	81
GOLD 4	6	2	32
Unknown	0	0	36
Smoking status
Current	59 (58.4)	7 (30.4)	517 (44.6)
Former	42 (41.6)	16 (69.6)	625 (53.9)
Never	0	0	17 (1.5)
Imaging biomarkers
Perc15 (HU)	−922 ± 27	−919 ± 30	−913 ± 30	−926 ± 30
VALD (HU)	82 ± 21	83 ± 23	87 ± 23	73 ± 23
TLC (liter)	5.9 ± 1.5	5.7 ± 1.5	5.3 ± 1.4	5.3 ± 1.4
Functional parameters
DLCO % predicted GLI	-	-	81.3 ± 20.6
FEV1% predicted	69.9 ± 25.3	69.4 ± 25.4	81.3 ± 22.9
FEV1/FVC	0.61 ± 0.15	0.61 ± 0.16	0.70 ± 0.13

Data are presented as the mean ± SD. Data in parentheses are percentages. C1 to C3 refer to data cohorts 1–3

COPDGene genetic epidemiology of chronic obstructive pulmonary disease, FD full-dose, RD reduced-dose, BMI body mass index calculated as weight in kilograms divided by height in meters squared, GOLD global initiative for obstructive lung disease, PRISm preserved ratio inspired spirometry, Perc15 15th percentile of the lung density histogram, VALD volume adjusted lung density, TLC total lung capacity, DLCO diffusing capacity of the lungs for carbon monoxide, GLI global lung function initiative, FEV1 forced expiratory volume in 1 s, FVC forced vital capacity

Virtual data analysis

Table S2 summarizes the quantitative assessment for model comparison. The standard 3D-UNet model showed superior performance, with an average Dice coefficient of 76.6% ± 8.8 and a bias ± limits of agreement of −0.5% ± 1.6. These results were achieved under optimal hyperparameters, including the ‘Adam’ optimizer and a 128 × 128 × 32 patch size. Figure 2 demonstrates a qualitative comparison of emphysema quantification using LAA-950 and AI-based methods using a virtual subject, scanned at two dose levels and reconstructed with two kernels. The emphysema masks show that the AI-based emphysema mask has a higher overlap with the groundtruth mask (Dice: 73.9–80.0%) and a more consistent quantification of emphysema (7.5–8.1%), as compared to the LAA-950 mask (Dice: 26.7–65.6%, emphysema quantification: 4.0–24.0%).

Fig. 2.

Qualitative demonstration of emphysema quantification across four renditions of CT scan (reduced-dose and full-dose, with soft and sharp kernel reconstructions) from a virtual patient, using (A–D) LAA-950 and (E–H) the AI-based model. I and J show ground truth renditions of CT scan and emphysema annotations, respectively. The percentage of emphysema varied by 20.0% and 0.6% using the LAA-950 and AI-based techniques, respectively

Clinical data analysis

Figure 3A–D shows a clinical scan at two dose levels and kernels, with the emphysema mask highlighted in red using LAA-950 (Fig. 3E–H) and the AI-based model (Fig. 3I–L). Emphysema quantification varied by < 18.3% with LAA-950 and < 2.0% with the AI-based model, indicating improved robustness of the AI-based approach.

Fig. 3.

A–D Representative CT images demonstrating qualitative demonstration of emphysema quantification for a 64-year-old male in the GOLD stage 2 category, imaged at two dose levels (reduced-dose and full-dose) and reconstructed with two kernels (soft and sharp). The percentage of emphysema varied by 18.3% and 2.0% using the (E–H) LAA-950 and (I–L) AI-based techniques, respectively

On the C1 cohort, AI-based masks achieved higher Dice scores (48.4% ± 24.2 vs 22.5% ± 20.2) and lower bias ± limits of agreement (1.0% ± 4.3 vs 2.1% ± 15.2). Stratified analysis further demonstrated consistently good segmentation performance across preserved ratio impaired spirometry (PRISm) and all GOLD stages (0–4), with mean Dice scores exceeding 40% in every category (Fig. S2 and Table S3). On the independent test set (C2), average Dice scores ranged from 60.3 to 64.8% across models trained on different folds. Figures S3A, B show axial slices with expert-annotated emphysema, and Fig. S3C–F show LAA-950 and AI-based masks, with the latter more closely resembling expert annotations shown in Fig. S3G, H. Lastly, agreement across imaging conditions was evaluated on the C3 cohort using Bland–Altman plots (Fig. 4), with the reference imaging setting defined as the full-dose, soft-kernel condition. The AI-based method approach reduced bias ± limits of agreement from 5.2% ± 8.0 (LAA-950) to 0.0% ± 3.6 and improved the reproducibility coefficient from 11.1% ± 0.2 (LAA-950) to 3.8% ± 0.2. Applying the AI model to harmonized CT data further improved bias ± limits of agreement to −0.1% ± 1.9 and reproducibility coefficient to 2.1% ± 0.1. Table 3 summarizes these metrics.

Fig. 4.

Bland–Altman plots comparing (A) LAA-950 and the artificial-intelligence (AI)-based emphysema quantification approach, applied to both (B) original and (C) harmonized CT data. Measurements were obtained from 1159 participants scanned under varying imaging conditions, including two dose levels and one/two reconstruction kernels. The reference condition was defined as the full-dose CT with soft-kernel reconstruction. The bias ± limits of agreement were improved from 5.2% ± 8.0 using the LAA-950 approach to 0.0% ± 3.6 and -0.1% ± 1.9 using the AI-based approach, applied on original and harmonized CT images, respectively

Table 3.

Comparison of LAA-950 and AI-based emphysema quantification across 101 COPDGene phase 1 participants, an independent dataset of 23 COPDGene phase 1 participants with expert annotations from scratch, and 1159 COPDGene phase 2 participants

	Approach	Dice (%)	Bias (%)	LOA (%)		RDC (%)
				Lower	Upper
C1 cohort	LAA-910	16.9 ± 22.7	23.0 ± 0.9	−2.8 ± 1.5	48.8 ± 1.5	15.1 ± 0.6
	LAA-950	22.5 ± 20.2	2.1 ± 0.5	−13.0 ± 0.9	17.3 ± 0.9	13.7 ± 0.7
	AI-based	48.4 ± 24.2	1.0 ± 0.2	−3.3 ± 0.3	5.3 ± 0.3	3.4 ± 0.5
C2 cohort	LAA-910	23.0 ± 21.3	18.6 ± 2.9	−9.1 ± 5.0	46.4 ± 5.0	45.5 ± 5.8
	LAA-950	32.7 ± 16.8	−0.9 ± 1.2	−12.0 ± 2.0	10.1 ± 2.0	10.7 ± 3.4
	AI—Fold 1	64.8 ± 9.3	−0.3 ± 0.2	−2.5 ± 0.4	1.9 ± 0.4	2.2 ± 0.4
	AI—Fold 2	62.3 ± 10.0	−0.1 ± 0.2	−2.0 ± 0.4	1.9 ± 0.4	1.9 ± 0.4
	AI—Fold 3	60.3 ± 11.3	−0.1 ± 0.2	−1.6 ± 0.3	1.3 ± 0.3	1.5 ± 0.3
	AI—Fold 4	60.4 ± 10.6	−0.3 ± 0.2	−2.2 ± 0.4	1.6 ± 0.4	2.0 ± 0.5
	AI—Fold 5	61.0 ± 10.3	0.6 ± 0.4	−3.5 ± 0.7	4.6 ± 0.7	4.0 ± 1.2
C3 cohort	LAA-910	−	5.0 ± 0.1	−5.6 ± 0.2	15.6 ± 0.2	12.8 ± 0.2
	LAA-950	−	5.2 ± 0.1	−2.7 ± 0.1	13.2 ± 0.1	11.1 ± 0.2
	AI at original CT	−	0.0 ± 0.0	−3.7 ± 0.1	3.6 ± 0.1	3.8 ± 0.2
	AI at harmonized CT	−	−0.1 ± 0.0	−2.0 ± 0.0	1.8 ± 0.0	2.1 ± 0.1

Data are mean ± standard deviation

Correlation with CT visual scores and pulmonary function tests

AI-based emphysema quantification showed stronger alignment with CT visual scores than LAA-950, with a higher Spearman correlation (0.77 vs 0.47). Kruskal–Wallis test indicated better differentiation across score categories for the AI-based method (H_AI = 128.6, H_LAA-950 = 64.1), and ordinal logistic regression showed a superior model, with higher log-likelihood (−255.7 vs −307.0), and lower AIC (523.5 vs 626.1) and BIC (543.7 vs 646.3). Figure 5 shows boxplots of emphysema measurements per CT visual score.

Fig. 5.

Box plots of (A) LAA-950 and (B) AI-based emphysema quantification across Fleischner visual emphysema scores, ranging from 0 to 5 (0 = absent, 1 = trace, 2 = mild, 3 = moderate, 4 = confluent, 5 = advanced destructive). The AI-based method showed stronger correlation with the visual scores (spearman correlation of 0.765 vs 0.473), greater group separation by Kruskal–Wallis test (H = 128.64 vs H = 64.12), and superior ordinal logistic regression performance, with higher log-likelihood (−255.75 vs −307.05) and lower AIC (523.50 vs 626.09) and BIC (543.67 vs 646.26)

Figure 6A–F demonstrate the relationship between pulmonary function tests and emphysema percentages from LAA-950 (A–C) and AI-based quantification (D–F). Univariate analysis revealed stronger nonlinear relationships for the AI method, along with higher R² values for DLCO and FEV1pp, and comparable performance for FEV1/FVC.

Fig. 6.

Scatter plots of (A–C) LAA-950 and (D–F) AI-based emphysema quantification against FEV1pp, FEV1/FVC, and DLCO. The AI-based method showed better nonlinear relationships across all pulmonary function tests, stronger linear correlations with DLCO and FEV1pp, and comparable performance for FEV1/FVC relative to LAA-950

Multicollinearity analysis and Pearson correlation matrix identified ‘WallAreaPct’ as highly collinear (VIF = 10.5 and Corr_Pi10 ¼ 0:86 and Corr_{AW T} ¼ 0:88; correlation matrix shown in Fig. S4). In multivariate analysis, the AI-based method improved predictive performance across all pulmonary function tests and imaging conditions when combined with airway disease biomarkers. Specifically, R² improved across imaging conditions from [0.21, 0.25] to [0.31, 0.32] for DLCO, [0.49, 0.52] to [0.52, 0.53] for FEV1pp, and [0.66, 0.68] to [0.69, 0.69] for FEV1/FVC when including the AI-based quantification, as compared to LAA-950. Consistent improvements were also observed in AIC and loglikelihood. Table 4 summarizes these metrics.

Table 4.

Statistical metrics summarizing the relationship between emphysema biomarkers (LAA-950 and AI-based quantification) and pulmonary function tests

	Type	Dose-Kernel	Pearson			Spearman			R2
			DLCO	FEV1pp	$\frac{FEV1}{FVC}$	DLCO	FEV1pp	$\frac{FEV1}{FVC}$	DLCO	FEV1pp	$\frac{FEV1}{FVC}$
Univariate Analysis	LAA-950	FDa - Soft	−0.40	−0.40	−0.64	−0.23	−0.19	−0.54	0.16	0.16	0.41
		RDa - Soft	−0.26	−0.28	−0.59	−0.14	−0.13	−0.50	0.07	0.08	0.35
		FD - Sharp	−0.39	−0.36	−0.64	−0.26	−0.20	−0.53	0.15	0.13	0.41
	AI-based	FD - Soft	−0.48	−0.42	−0.60	−0.43	−0.29	−0.62	0.23	0.18	0.36
		RD - Soft	−0.48	−0.44	−0.62	−0.41	−0.27	−0.63	0.23	0.19	0.39
		FD - Sharp	−0.50	−0.43	−0.62	−0.47	−0.30	−0.63	0.25	0.18	0.38
			AIC			Log Likelihood			R2
	Type	Dose-Kernel	DLCO	FEV1pp	$\frac{FEV1}{FVC}$	DLCO	FEV1pp	$\frac{FEV1}{FVC}$	DLCO	FEV1pp	$\frac{FEV1}{FVC}$
Multivariate Analysis	LAA-950	FD - Soft	7716	7445	−2209	−3853	−3717	1109	0.25	0.52	0.68
		RD - Soft	7737	7467	−2147	−3863	−3728	1078	0.21	0.49	0.66
		FD - Sharp	7723	7461	−2191	−3856	−3725	1100	0.24	0.50	0.68
	AI-based	FD - Soft	7642	7427	−2217	−3816	−3708	1113	0.31	0.52	0.69
		RD - Soft	7620	7400	−2219	−3805	−3695	1114	0.31	0.53	0.69
		FD - Sharp	7627	7423	−2228	−3808	−3706	1119	0.32	0.52	0.69

Results are shown for both univariate and multivariate analyses. Multivariate models included either LAA-950 or AI-based quantification along with imaging biomarkers of small airway disease and airway remodeling

^aFD and RD refer to full-dose and reduced-dose scans

Discussion

This study demonstrated the variability of LAA-950 across imaging conditions and proposed a deep learning segmentation model as a promising alternative for emphysema quantification. To develop and validate the framework, we leveraged both virtual and clinical datasets. The virtual dataset enabled model pretraining with ground-truth emphysema masks, achieving a Dice coefficient of 76.6% ± 8.8. The clinical dataset allowed for fine-tuning via a human-in-the-loop strategy, yielding a Dice of 48.4% ± 24.2, outperforming LAA-950 (Dice: 22.5% ± 20.2). Similar improvements were observed on an independent test set (average Dice_LAA–950 range: 60.3–64.8% vs Dice_LAA–950: 32.7% ± 16.8). Further assessment of emphysema quantification robustness showed minimal variation (bias ± limits: 0.0% ± 3.6) compared with LAA-950 (5.2% ± 8.0) across 1159 cases with varying dose and reconstruction kernels. In evaluating clinical relevance, AI-based quantification showed a stronger correlation with Fleischner visual emphysema scores than LAA-950 (Spearman correlation: 0.77 vs 0.47). Consistent with prior studies [37, 38], higher emphysema measures corresponded to increasing visual grades. Greater variability at higher grades may be attributable to the limited number of data points in these categories (10 and 2 cases out of 101 for grades 4 and 5, respectively). The AI method also demonstrated stronger associations with pulmonary function tests, with higher R², log-likelihood, and lower AIC across imaging conditions in both univariate and multivariate models, supporting its superior accuracy, robustness, and clinical utility.

This study evaluated segmentation performance across different datasets and reported variable Dice values. This variation reflects the distinct annotation and evaluation methodologies used for each dataset, precluding direct comparison of the reported Dice scores. The virtual dataset used ground-truth 3D emphysema masks with homogeneous density per case. In contrast, accurate emphysema annotation is inherently challenging on clinical datasets (C1 and C2) given the heterogeneous and diffuse nature of the disease [39–42]. Furthermore, C1 was evaluated using volumetric 3D annotations, whereas C2 was evaluated using 20 uniformly sampled axial slices, resulting in inherently different Dice measurements. Lastly, C1 was used for model fine-tuning within the human-in-the-loop framework; hence, its score may be subject to anchoring bias.

Previous studies have developed deep learning models for emphysema and air trapping segmentation on chest CT. Ram et al [43] applied a feature-based PRM method using invariant scattering convolutional networks to quantify fSAD and emphysema, showing strong agreement with LAA-950 (R² = 0:98), and improved noise robustness—consistent with our findings. Another study adapted a deep learning model from an HMMF framework [15], reporting Dice scores of 70.2% ± 10.9 (in-distribution) and 65.8% ± 14.7 (out-of-distribution). While our segmentation performance was comparable, we further examined robustness across imaging conditions and correlations with pulmonary function tests. Related work on air trapping quantification [44, 45] also showed improved agreement with radiologist assessments and greater robustness to suboptimal exhalation. Building on these efforts, our study systematically evaluated emphysema quantification variability across diverse imaging settings, comparing AI-based results with LAA-950 thresholding.

Our findings have important clinical implications. In a cohort of 1159 COPDGene participants scanned under varying doses and kernels, LAA-950 exhibited a reproducibility coefficient of 11.1% ± 0.2—substantial, given that a 2% annual change is often the minimum detectable difference in clinical trial design [1]. The AI-based method reduced the reproducibility 2.9-fold, to 3.8% ± 0.2, closely aligning with the weighted mean reproducibility of 3.7% ± 0.7 reported by QIBA [46] from meta-analyses of reproducibility studies. Applying the AI-based model to harmonized CT data [10] further reduced it to 2.1% ± 0.1—a 1.8-fold improvement over the non-harmonized version. These results highlight the value of standardized, low-variability inputs in enhancing the robustness of emphysema quantification across dose levels and reconstruction kernels.

This study had some limitations. First, the AI-based model was trained on a limited number of virtual and clinical subjects. While expanding clinical datasets requires extensive manual annotations, incorporating additional virtual subjects is more feasible, particularly given advances in digital twin generation [47]. Second, clinical annotations were derived by a single reader; hence, inter-reader variability could not be assessed. Third, testing on the C2 cohort used uniformly sampled slices rather than full lung volume. Fourth, the human-in-the-loop training strategy may introduce anchoring bias, mitigated through iterative multi-planar review by the reader with and without model-generated masks. Fifth, the virtual dataset was limited to a single vendor; hence, the reproducibility and robustness findings should be interpreted in this context. Nevertheless, multi-vendor clinical data demonstrated generalizability across three vendors. Future studies may further validate these findings using multi-vendor virtual datasets. Sixth, downsampling along the z-axis may reduce depiction of very small structures, although its impact on patient-level emphysema quantification is expected to be minimal. Finally, this study was not designed for longitudinal evaluation or model optimization, and future work should assess disease progression, improved model performance, and computational efficiency for real-time use.

In conclusion, this study introduced a deep learning framework for emphysema quantification as a promising alternative to the conventional LAA-950 thresholding approach. Pretrained on virtual data and fine-tuned with clinical scans, the model showed improved agreement with radiologist assessments, stronger correlations with pulmonary function tests, and improved reproducibility across imaging conditions. This approach shows strong potential for reliable emphysema quantification, particularly in light of ongoing efforts toward lower radiation doses. Future studies may explore the longitudinal utility of this method for disease monitoring and the clinical relevance of phenotypic features derived from emphysema masks.

Supplementary Material

The online version contains supplementary material available at https://doi.org/10.1007/s00330-026-12800-4 .

Key Points.

Question

Can deep learning-based emphysema segmentation on CT improve the accuracy and robustness of emphysema quantification relative to the traditional LAA-950 biomarker?

Findings

The deep learning model outperformed LAA-950 in terms of accuracy and reproducibility and demonstrated stronger correlations with visual emphysema scores and pulmonary function tests.

Clinical relevance

The improved performance of the deep learning approach enables more reliable emphysema quantification across imaging conditions, with closer agreement to radiologist assessments and pulmonary function tests.

Abbreviations

COPDGene

Genetic Epidemiology of COPD

DLCO

Diffusing capacity of the lungs for carbon monoxide

FEV1

Forced expiratory volume in 1 second

FVC

Forced vital capacity

FEV1pp

Percent-predicted forced expiratory volume in 1 second

FD

Full dose

LAA-950

Low-attenuation area below −950 HU