Virtual Imaging-Guided Harmonization of Chest X-rays Using a Generative Adversarial Neural Network

Abstract

Objective:

To reduce variability in chest radiography (CXR) from acquisition and post-processing, and assess whether harmonization improves image quality and downstream diagnostic performance.

Methods:

A generative adversarial network (GAN) was trained exclusively on virtual images produced by a Virtual Imaging Trial (VIT). The model maps non-harmonized CXRs to noise-free, unprocessed references using a dual U-Net with contrastive loss. Evaluation spanned virtual, physical-phantom, and clinical data using generic (GIQMs) and chest-specific (CIQMs) metrics. A phantom benchmark compared the method with ComBat harmonization and a conventional denoising algorithm. Clinical generalization was tested on VinDr-CXRs, examining feature compactness via Uniform Manifold Approximation and Projection (UMAP). We also assessed NIH-CXRs, quantifying diagnostic accuracy with bootstrap uncertainty.

Results:

Compared with non-harmonized images, harmonized CXRs yielded approximately 89% lower NRMSE, 50% higher PSNR, and 86% higher SSIM. On phantom data, CIQM variability fell from 0.255 to 0.026 and was reduced more consistently than with ComBat or denoising. Clinical analyses on VinDr-CXRs showed tighter UMAP clusters for harmonized features, indicating suppression of acquisition-related variability. On NIH-CXRs, training and testing a classifier in the harmonized domain improved diagnostic accuracy over the non-harmonized domain and lowered cross-domain sensitivity; pleural and cardiac categories showed consistent gains, while texture-dependent labels exhibited task-dependent effects.

Conclusion:

VIT-guided training enables a GAN to harmonize CXRs toward a raw reference, improving quantification variability and harmonizing image representations across systems.

Significance:

The proposed virtual-to-clinical strategy is scalable and generalizable, offering a practical path to standardized CXR appearance and reliable downstream detection across institutions.

I. Introduction

MEDICAL imaging plays a critical role in clinical decision-making, yet image variability remains a pervasive challenge. Differences in scanner hardware, acquisition protocols, post-processing techniques, and patient-specific factors lead to inconsistencies in how anatomy and pathology are rendered. In chest radiography (CXR), such variability is especially pronounced due to inconsistent post-processing across institutions [1], [2]. These differences affect both perceived image quality and quantitative analysis, potentially compromising diagnostic accuracy and downstream AI applications [3–5].

Various methods have been proposed to address harmonization challenges. These include domain adaptation [6], histogram equalization [7], wavelet-based denoising [8], ComBat [9], and out-of-distribution detection [10]. More recent efforts leverage deep learning models, such as diffusion models [11], transformers [12], and generative adversarial networks (GANs) [13]. While these approaches offer promise, their performance is often constrained by the need for large, diverse, and well-curated training datasets [14]. To compensate for clinical data limitations, researchers have used physical phantoms [15], [16]; however, their restricted anatomical and pathological variability limits generalization.

Other studies focus on leveraging intrinsic image characteristics through multi-task learning [17], physics-informed modeling [18], [19], or contrastive learning [20]. These methods depend on defining accurate similarity metrics and often require auxiliary information, which can be difficult to obtain and may introduce bias if improperly formulated [21].

In the context of CXR harmonization, practical limitations such as limited training data, expert labeling cost, and high image dimensionality make model training challenging [22], [23]. While patch-based training [24], [25] increases data volume, it sacrifices global context and may introduce artifacts [26]. Additional strategies like transfer learning, data filtering, and synthetic augmentation [27] help improve robustness but still rely on high-quality clinical datasets.

To address these challenges, we propose a harmonization approach built entirely within a Virtual Imaging Trial (VIT) framework. VITs provide access to diverse, clinically realistic digital phantoms [28] and simulate commercial imaging systems [29], enabling controlled acquisition and access to noise-free ground truth images. We leverage these advantages to develop a DNN-based model that harmonizes CXRs to a standardized raw image target. The model is trained using contrastive learning to enforce anatomical consistency across renditions of the same subject and evaluated across virtual, phantom, and clinical datasets to assess generalizability and performance. Prior studies have emphasized image-quality metrics, but the practical value of harmonization is determined by its effect on downstream clinical tasks. Accordingly, we broaden our assessment to large-scale clinical benchmarks, quantifying how a harmonizer trained in the virtual domain influences multi-label diagnostic performance across domains.

II. Methods

We developed a GAN harmonizer trained on diverse virtual CXRs to map non-harmonized images to a standardized, raw-like target—defined as unprocessed, sharp, and noiseless—to jointly address restoration and denoising. Two sequentially trained U-Nets (ReversalUNet and RestoralUNet) were employed to achieve this mapping using quasi-ground truth (QGT, containing realistic noise but no post-processing) and ground truth (GT) images. While initial variability analysis was conducted across post-processing software, subsequent evaluations were performed using a fixed post-processing configuration (software “A”) to isolate the effects of harmonization. Performance was assessed using generic (GIQMs) and clinical (CIQMs) image quality metrics [30]. Results were benchmarked against both a composite denoising method—combining wavelet shrinkage [31] with adaptive histogram equalization [32], and a standard U-Net model. Generalizability was assessed in three ways:

1. on a physical phantom, by measuring CIQM variability after denoising and harmonization (via ComBat [9] and our proposed method);

2. on a clinical dataset, by analyzing latent feature distributions with UMAP [33] and

3. on a second clinical dataset, by evaluating a downstream diagnostic task to demonstrate generalization in end-to-end performance.

A. Data Generation

213 digital XCAT phantoms [28], [34–36], including both healthy and diseased models [37], were simulated using DukeSim [38] at four dose levels (3, 9, 14, and 20 milliampere seconds) with a pixel size of 0.144 mm. The virtual cohort comprised 213 XCAT digital human models spanning both healthy and diseased anatomy: 20 healthy, 60 COVID-19, 67 COPD (including emphysema), and 66 lung cancer. The cohort was approximately sex-balanced with 106 male models (age 58.4±12.4 years; BMI 28.0±5.4) and 107 female models (age 58.8±11.2 years; BMI 28.2±6.8). External devices (e.g., endotracheal tubes, ECG leads) were not simulated. Modeling choices followed validated XCAT/DukeSim practices to ensure anatomically and system-accurate image formation. The simulations were performed to mimic image acquisition at 120 kV and produced two sets of images: ground truth (GT, ideal) and quasi-ground truth (QGT, containing realistic noise but no post-processing).

Commercial post-processing software “A” and “B” with 10 parameter variations each produced 80 non-harmonized CXRs per phantom. Post-processing parameters were selected to replicate real-world clinical imaging conditions.

B. Data Pre-processing

Non-harmonized CXRs were converted to 16-bit and then were scaled to [0,1]. GTs and QGTs were transformed by Box-Cox method [39] to reduce skewness. All CXRs, GTs, and QGTs were padded to 3000×3000, then downsampled to 384×384. One phantom’s data (80 images) was held out for testing; the rest were split into training (13,568) and validation (3,392).

C. CXR Harmonizer Training

1). Networks Architecture:

We used a Pix2Pix framework [40], in which the generator consisted of two U-Nets—ReversalUNet and RestoralUNet—each incorporating spatial and channel-wise attention mechanisms [41]. The outputs of these two networks were merged through a long-range spatial attention module to produce the final output of the generator. The harmonizer model took a non-harmonized CXR as input and generated a harmonized raw image as output. The discriminator comprised three convolutional layers with leaky ReLU activations.

2). Networks Weight Initialization:

Each U-Net was pretrained (100 epochs) using ℒ₁ loss—ReversalUNet on non-harmonized/QGT pairs, and RestoralUNet on QGT/GT pairs—before adversarial training. This separation improved learning stability and convergence.

3). Loss Functions:

Discriminator loss used LSGAN [42]:

$${ℒ}_{\text{D}}(G,D)=\frac{{\mathbb{E}}_{I_{GT}}}{2}\left[{\left(D\left(I_{GT}\right)-1\right)}^2\right]+\frac{{\mathbb{E}}_{I_{nh}}}{2}\left[{\left(D\left(G\left(I_p\right)\right)\right)}^2\right]$$ (1)

$${ℒ}_{\text{LSGAN}}\left(I_h\right)={\mathbb{E}}_{I_h}\left[{\left(D\left(I_h\right)-1\right)}^2\right]$$ (2)

To enforce consistency across system variations, we introduced a novel perceptual contrastive loss function tailored for harmonization, inspired by the principles of contrastive learning [43]. The general concept leverages similarity across different views of the same subject while maximizing dissimilarity across different subjects. In our case, positive samples were CXRs of the same virtual patient acquired under varying imaging conditions, and negative samples were selected from different virtual patients. System-variant augmentations (e.g., signal noise or jitter) were used to generate positive pairs, whereas both system- and geometry-variant augmentations were used for negative pairs.

Feature vectors from harmonized samples ($f_j^{+},f_j^{2+}$, and $f_j^{-}$) were extracted from the downsampling layers (N = 5) of a pretrained VGG19 [44] and were used to compute the following custom contrastive loss:

$${ℒ}_{\text{C}}\left(f^{+},f^{2+},f^{-}\right)=1-\text{cos}\left(f,f^{+}\right)+\text{max}\left(0,\text{cos}\left(f^{+},f^{-}\right)\right)+\text{max}\left(0,\text{cos}\left(f^{2+},f^{-}\right)\right)$$ (3)

The perceptual contrastive loss was then formulated as:

$${ℒ}_{\text{PCL}}={\displaystyle \sum }_{j=1}^N{ℒ}_{\text{C}}\left(f_j^{+},f_j^{2+},f_j^{-}\right)$$ (4)

The total generator loss then was:

$${ℒ}_{\text{G}}={ℒ}_{\text{LSGAN}}+{\lambda }_{L_1}\Vert I_{GT}-I_h\Vert +{\lambda }_{PCL}{ℒ}_{\text{PCL}}$$ (5)

with λ_L1 = 100 and λ_PCL = 1.

D. Ablation Study

To quantitatively evaluate the individual contributions of the proposed architectural and training enhancements, we conducted an ablation study. Specifically, we implemented a baseline U-Net model following the conventional Pix2Pix architecture. This model excluded the perceptual contrastive loss $\left({ℒ}_{\text{PCL}}\right)$ and the dual-stage U-Net design relying only on standard GAN loss components ($\left({ℒ}_1\right)$ and the LSGAN losses). To ensure a fair comparison, the baseline model was trained using the exact same training and validation datasets, batch size, optimizer, learning rate schedule, and number of training epochs as the proposed method. Post-processing of the output images from the baseline model was also kept identical—using software “A” with default parameters—ensuring that any observed differences in performance could be attributed to model design rather than post-processing artifacts. By comparing the baseline model’s outputs to those of the proposed harmonization pipeline across these standardized criteria, we aimed to isolate and quantify the performance gains attributable to the dual-stage U-Net configuration and the incorporation of the perceptual contrastive loss function.

E. CXR Harmonizer Performance Evaluation

Evaluation of the harmonized images was performed using both GIQMs and CIQMs. The GIQMs included normalized root-mean-square error (NRMSE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM), providing a general assessment of image fidelity. The CIQMs [30] consisted of ten task-specific radiological features relevant to diagnostic quality, including lung grey level (GL), lung detail (LD), lung noise (LN), rib-lung contrast (RC), rib sharpness (RS), mediastinum detail (MD), mediastinum noise (MN), mediastinum alignment (MA), sub-diaphragm contrast (SC), and rib resolution (RR). The benchmark denoising approach combined VisuShrink wavelet denoising [31] with adaptive histogram equalization [32].

1). Virtual CXR Image Analysis:

GIQMs and CIQMs were calculated on harmonized, U-Net, and denoised images. Statistical comparisons between harmonized and non-harmonized images were conducted using paired t-tests with Bonferroni correction.

2). Anthropomorphic Phantom’s CXR Images Analysis:

A physical phantom [45] was imaged using software “A” with 5 different protocols. Harmonized outputs were post-processed with default parameters. CIQMs’ coefficient of variation (CV) was compared across harmonized, non-harmonized, and denoised CXRs. To benchmark our harmonizer against a conventional statistical harmonization approach, we additionally applied ComBat [9] to the same phantom dataset. ComBat has been widely used to address batch effects in radiomic features but is limited by assumptions of linear shifts and the need for multiple samples per batch. For comparison, the non-harmonized features were harmonized using ComBat with the first acquisition defined as the reference batch, and results were evaluated using the same CIQMs.

3). Clinical CXR Images Analysis:

a). VinDr-CXR Validation:

3000 images from the VinDr-CXR dataset [46] were harmonized and then post-processed using software “A”. This demonstrated the harmonizer’s ability to generalize to unseen post-processing, confirming its robustness in real-world imaging scenarios. Since clinical datasets lack ground truth or multiple renditions of the same patient, direct harmonization accuracy could not be measured. Therefore, evaluation focused on statistical and representation-level consistency using two proxy metrics: histogram entropy and dynamic range. These metrics were chosen because they reflect texture complexity and intensity contrast—key indicators of CXR quality and clinical interpretability.

To further evaluate the consistency and variability of image representations before and after harmonization, Uniform Manifold Approximation and Projection (UMAP) [33] was employed. UMAP is a non-linear dimensionality reduction technique that preserves both local and global structure from high-dimensional feature spaces in a lower-dimensional manifold. This property makes it particularly well-suited for visualizing latent similarities and differences among medical images, where feature relationships are complex and often non-Euclidean.

Using UMAP, deep feature representations of harmonized and non-harmonized CXRs were projected into a low-dimensional space. The spatial arrangement and compactness of the resulting clusters were then analyzed to evaluate the extent to which harmonization reduced variability stemming from acquisition and post-processing differences. A tighter clustering of harmonized images suggests that system-induced variations were successfully minimized, allowing patient-specific characteristics to emerge as the primary source of variation. Therefore, UMAP was used to visualize and validate how harmonization structured the latent feature space of clinical CXRs.

b). Downstream Diagnostic Evaluation on NIH ChestX-ray14:

We evaluated the clinical utility of the proposed harmonization using a multi-label disease detection task on NIH ChestX-ray14 [47]. The dataset was partitioned into 35k/5k/5k train/validation/test images at the subject level. For each split we prepared two domains: Non-harmonized (NH) images (standard vendor/post-processed) and their corresponding Harmonized (H) counterparts generated by our proposed algorithm. Identical classifiers (DenseNet-121) were trained from scratch in each domain with the same initialization, optimization, and regularization settings. Training proceeded for 100 epochs, with the best checkpoint selected by validation Area Under the Receiver Operating Characteristic (AUROC) curve; early overfitting was observed around 70 epochs for Non-harmonized and 90 epochs for Harmonized domains. Evaluation followed a 2x2 design (train→test: NH→NH, Harmonized→H, NH→H, H→NH). We report per-pathology AUROC and Area Under the Precision–Recall Curve (AUPRC) along with averages. Uncertainty was quantified by nonparametric bootstrap resampling over test cases to obtain mean±SD for each label and condition.

III. Results

A. Virtual CXR Image Analysis

Quantitative results in Figure 3 indicate substantial improvements in image quality metrics for harmonized CXRs compared to non-harmonized ones. Specifically, the harmonized images achieved an approximate 89% reduction in NRMSE, 50% improvement in PSNR, and 86% increase in SSIM relative to non-harmonized images. In addition to these gains, the harmonized images also exhibited significantly lower variability in these metrics.

Fig. 3.

Comparison of generic image quality metrics—NRMSE, PSNR, and SSIM—across non-harmonized, harmonized (using the proposed algorithm and U-Net model), and denoised CXRs, post-processed with two distinct software tools. PSNR values were normalized for comparability by using the formula PSNR/(100 dB)×100. Lower NRMSE values indicate better quality, while higher PSNR and SSIM values correspond to enhanced image quality. Regardless of post-processing method, harmonized CXRs consistently showed reduced variability and more accurate alignment with processed ground truth (GT) images compared to non-harmonized and denoised CXRs.

Compared to the conventional denoising and UNet-based methods, the proposed harmonization approach outperformed both alternatives in terms of average metric improvement and consistency. While the UNet model yielded similar reductions in variability, it fell short in achieving the same level of improvement in absolute metric values. Statistical tests in Table II further confirmed the superiority of the proposed method, with all improvements reaching significance (corrected p-value << 0.01).

TABLE II.

Comparison of CIQMs across all methods. Bold t values indicate statistically significant differences (p < 10⁻²) between non-harmonized (NH) and harmonized (H) images. For H, U, and D, the coefficient of variation (CV = SD / Mean) is reported in parentheses; the lowest CV per metric is highlighted in bold.

Metric	NH	GT	H (CV)	U (CV)	D (CV)	t (NH vs H)	Eff. Sz (NH vs H)
GL	1522±255	1952	1970±22 (0.011)	1699±27 (0.016)	1523±257 (0.169)	11.53	2.35
LD	43.1±14	49	42.1±3.3 (0.078)	40.7±3.3 (0.081)	47.9±14 (0.292)	4.11	0.84
LN	71±33	32.5	11.5±1.5 (0.130)	8.1±0.8 (0.099)	74±32 (0.432)	3.79	0.77
RC	0.21±0.10	0.15	0.17±0.00 (0.000)	0.16±0.01 (0.063)	0.21±0.10 (0.476)	4.39	0.90
RS	89.7±0.07	89.8	89.7±0.02 (0.0002)	89.7±0.03 (0.0003)	89.8±0.05 (0.0006)	4.11	0.84
MD	9.4±2.4	14.5	11.7±1.0 (0.085)	10.1±1.1 (0.109)	9.7±2.2 (0.227)	3.99	0.82
MN	65±31	20.9	12.7±0.5 (0.039)	11.2±0.6 (0.054)	68±32 (0.471)	7.95	1.62
MA	7.6±2.5	4.4	5.9±1.9 (0.322)	7.4±0.7 (0.095)	6.7±2.0 (0.299)	5.55	1.13
SC	−0.19±0.30	−0.25	−0.19±0.30 (1.579)	−0.27±0.20 (0.741)	−0.15±0.32 (2.133)	3.97	0.81
RR	0.36±0.05	0.39	0.38±0.04 (0.105)	0.40±0.04 (0.100)	0.36±0.04 (0.111)	1.74	0.36

Figure 4 and Table II show that harmonized images align more closely with the ground truth (GT) in CIQMs, demonstrating reduced variability and bias—except for rib sharpness and mediastinum detail, where deviations were likely due to training on downsampled data.

Fig. 4.

Chest-specific quality metrics variability and bias with respect to the GT rendition in the virtual CXRs. Across all metrics, harmonized images show lower NRMSE and variability compared to non-harmonized images.

B. Anthropomorphic Phantom’s CXR Images Analysis

Figure 5 visually compares non-harmonized, harmonized, and denoised phantom CXRs. Harmonized images showed more consistent appearance across varying post-processing conditions. Figure 6 reports coefficients of variation (CVs), and Table III summarizes mean±SD values for CIQMs across methods, including ComBat (CB). Compared to NH and CB, H consistently reduced variability across most CIQMs, with the most pronounced improvements in GL, LN, RC, MN, and SC. While CB lowered variability relative to NH in some metrics (e.g., MD and MN), its residual SDs remained markedly larger than those observed for H. These results highlight that the proposed harmonizer provided superior stabilization of CIQMs across acquisitions compared to both denoising and CB.

Fig. 5.

Non-harmonized renditions of the physical phantoms (top row), harmonized CXRs (middle row), and denoised CXRs (bottom row). Compared to the non-harmonized CXRs, the corresponding harmonized and denoised CXRs represent more consistent appearance.

Fig. 6.

Coefficient of variation of CIQMs for the non-harmonized CXRs, ComBat harmonization, harmonized CXRs, and denoised CXRs for the anthropomorphic physical phantom.

TABLE III.

CIQMs over physical phantom CXRs

	NH	CB	H	D
GL	985±444	939±223	1255±11	373±25
LD	57±5	57±2	53±4	65±8
LN	44±7	42±2	14±1	38±3
RC	0.42±0.24	0.39±0.11	0.27±0.01	0.81±0.08
RS	89.86±0.04	89.85±0.03	89.80±0.02	89.88±0.02
MD	16.4±3.3	18.1±2.4	25±1.6	34±1.0
MN	69±13	74.8±9.8	20±0.4	73±8
MA	12±0.18	11.6±0.03	10±0.17	12±0.44
SC	1.07±0.39	1.05±0.20	0.93±0.01	1.65±0.02
RR	0.54±0.11	0.51±0.06	0.43±0.02	0.56±0.06

NH = non-harmonized, CB = ComBat, H = harmonized, D = denoised.

C. Clinical CXR Images Analysis

1). VinDr-CXR Validation:

Out of 3000 harmonized CXRs, only 42 cases (⩽ 2%) were identified as potential harmonization failures. This initial screening was based on a SSIM threshold of ¡0.9 between each non-harmonized image and its corresponding post-processed harmonized version. The threshold served as a conservative indicator of substantial structural deviation. Flagged cases were then visually reviewed, and only those showing clear anatomical inconsistencies—such as missing or deformed lung or rib structures—were classified as harmonization failures.

Figure 7 illustrates that harmonized images exhibited more consistent histogram entropy and dynamic range compared to their non-harmonized counterparts. Specifically, the dynamic range decreased while entropy increased. This suggests that the harmonization process suppressed outlier intensities and artifacts, leading to a more compact but informative intensity distribution. The increase in entropy reflects a more efficient use of the available intensity values, enhancing contrast in clinically relevant regions while reducing unnecessary variability.

Fig. 7.

Histogram entropy and dynamic range for both harmonized and non-harmonized CXRs in VinDr-CXR dataset. The harmonized CXRs being post-processed using the default parameters of software A. While the dynamic range was compressed, the rise in histogram entropy suggests a more balanced distribution of pixel intensities, possibly enhancing local contrast despite the overall range reduction.

Figures 8a–8b show UMAP projections of deep feature representations for both harmonized and non-harmonized CXRs. In the UMAP_non–harm space (Figure 8a), non-harmonized images appear widely dispersed, reflecting variability from acquisition and post-processing differences. After harmonization, the projections form tighter clusters, indicating that system-induced variability was effectively suppressed. In the UMAP_harm projection (Figure 8b), harmonized and non-harmonized images share a more balanced feature space, where patient-specific attributes—rather than acquisition differences—become the dominant factors shaping the latent representation. This shift confirms that the harmonization process successfully isolates anatomical and pathological features from technical confounders. Qualitative improvements are further illustrated in Figure 9.

Fig. 8.

The low dimensional domains obtained from UMAP_{non_harm}(a), and UMAP_harm (b) on VinDr-CXR dataset. The occupied area with harmonized and non-harmonized feature space are depicted in blue and red shades, respectively. In (a), post-processing variability was identified as the primary factor of influence. Consequently, harmonized images, generated under a single, specific post-processing condition, were projected as a closely grouped cluster. The images shown in pairs illustrate both non-harmonized and harmonized examples derived from the nearest samples to the center of each cluster. In (b), the post-processing effects were separated from the latent features, resulting in patient attributes becoming the primary influencing factors. As a result, both harmonized and non-harmonized images occupied a projection domain of approximately equal size.

Fig. 9.

Examples of the harmonized (in blue squares) and corresponding non-harmonized (in red squares) images from VinDr-CXR dataset.

2). Downstream Diagnostic Evaluation on NIH ChestX-ray14:

Figure 10 presents per-pathology AUROC as mean±SD from bootstrap resampling for all four train→test conditions, with positive case counts indicated beside each label. Two consistent patterns emerge. First, in-domain evaluation outperforms cross-domain evaluation, indicating sensitivity to a train–test domain shift. Second, training and testing in the Harmonized domain yields a modest but reproducible improvement over the Non-harmonized domain for several findings (notably pleural and cardiac categories), while a small subset of texture-dependent or rarer findings shows neutral or slightly lower performance. The error bars are generally comparable or narrower for H→H relative to NH→NH, suggesting reduced variability for most labels. Cross-domain conditions, particularly H→NH, exhibit larger performance gaps, reinforcing the benefit of aligning training and deployment image statistics via harmonization.

Fig. 10.

Per-pathology diagnostic performance on NIH ChestX-ray14 under four train→test conditions. Points show AUROC (mean) with bootstrap SD error bars (B=500 resamples) for each label; x-axis annotations include the number of test cases (N). Color coding: NH→NH (red), H→H (blue), NH→H (dark gray), H→NH (dark green).

IV. DISCUSSION

This study leveraged VITs to develop and train a deep neural network for harmonizing CXR images toward a standardized, raw imaging space. VITs provided access to anatomically realistic digital phantoms, customizable acquisition settings, and noise-free ground truth references—resources that are otherwise difficult to obtain in clinical settings. This controlled environment enabled not only the creation of a diverse and balanced training dataset, but also the definition of a perceptual contrastive loss function specifically tailored to enforce consistency across patient-specific but system-varied image renditions.

The results demonstrated significant improvements in both GIQMs and CIQMs. Compared to conventional approaches such as ComBat and traditional denoising algorithms, the proposed harmonization method reduced variability and bias more effectively, preserving essential anatomical and contrast features while enabling downstream post-processing to be applied uniformly. While ComBat and denoising each achieved partial reduction in variability, residual dispersion across several CIQMs remained higher than with the proposed harmonizer. These enhancements are particularly valuable in multi-institutional or longitudinal studies where variability in acquisition and post-processing can confound both clinical interpretation and AI-based analysis.

The low-dimensional projections derived via UMAP offered key insight into the latent effects of harmonization. In the non-harmonized domain, UMAP revealed a dispersed and inconsistent feature space, where post-processing differences dominated the structure. In contrast, harmonized images clustered more compactly, suggesting that the model successfully disentangled system-level variability from patient-specific features. This shift from acquisition-dependent to anatomy-dependent variance validates the harmonization objective and confirms the model’s ability to improve consistency in latent representations. Moreover, the more uniform harmonized manifold reflects improved feature alignment, which is critical for both radiological interpretation and downstream machine learning applications. These findings highlight the utility of UMAP not just as a visualization tool, but as a proxy for latent structure analysis in the absence of ground truth clinical labels.

The ablation study further confirmed the added value of the proposed method over a baseline U-Net. While the U-Net achieved reductions in metric variability, it underperformed in replicating clinically important CIQMs—particularly in metrics like lung grey level and mediastinum detail. This reinforces the importance of incorporating contrastive learning and multi-stage architecture tailored for harmonization tasks, rather than relying solely on conventional generative models.

The NIH14 experiment shows that harmonization yields downstream diagnostic gains, improving average discrimination and reducing domain sensitivity. Improvements concentrate in findings whose cues are robust to contrast re-mapping and noise suppression (e.g., pleural and cardiac), while small declines for Emphysema, Infiltration, and Hernia likely reflect attenuated parenchymal texture, and label scarcity. Cross-domain gaps further emphasize the need to align training and deployment image statistics; harmonization provides a principled route to that alignment while preserving clinically relevant structure.

Although the harmonizer was trained on 213 XCAT phantoms that represent healthy subjects and a limited set of disease models, the NIH ChestX-ray14 dataset encompasses a broader spectrum of pathologies and institutional imaging practices. As a result, our method may not eliminate all aspects of dataset shift. The training objective and the contrastive design target acquisition and post processing variability rather than pathology labels, so the learned mapping is best interpreted as a correction for system induced covariate shift. This interpretation is supported by the UMAP analysis, where harmonization reduces feature space structure related to post processing, and by the downstream results in which gains are largest for findings driven by global density and contour cues, while texture dependent findings show smaller or mixed effects. We therefore view the current model as addressing one major component of dataset shift, with residual differences due to population and disease mix remaining. Extending the virtual cohort and incorporating clinical adaptation strategies will be important directions to further improve robustness across diverse patient populations

Despite promising results, the approach has limitations. Generalization to newer systems or post-processing pipelines may require retraining; however, this is relatively scalable with VIT because new variations can be simulated more rapidly and comprehensively than they can be collected clinically. The proposed architecture benefits from access to ground truth and controlled variation, which supports faster convergence in contrastive learning. The training was done on downsampled images due to computing constraints. This may degrade resolution-sensitive metrics and contribute to sub-optimal performance on texture-dependent findings such as emphysema and infiltration, where parenchymal texture cues are attenuated by downsampling. However, our validation confirmed that key anatomical and contrast features were preserved, and diagnostic performance improved despite this resolution reduction, suggesting that harmonization benefits are driven primarily by global consistency rather than pixel-level detail. In addition, the uniform post-processing used to render harmonized images may be sub-optimal across different clinical tasks, as tasks emphasize distinct image cues (e.g., parenchymal texture vs. silhouette/contour). As a direction for future work, we propose task-aware post-processing: for each downstream task, select and fix a parameter set for post-processing software that maximizes task accuracy. Future studies should also explore full-resolution training and broader phantom diversity. While the dual U-Net structure proved effective, more compact architectures could be assessed as data diversity increases.

Overall, this work demonstrates the feasibility and efficacy of using VITs for training harmonization algorithms, and shows that harmonizing CXRs to a raw reference domain can meaningfully reduce acquisition-related variability. By improving both pixel-level quality and featurespace consistency, the proposed approach lays a foundation for more reliable and standardized CXR interpretation across institutions and imaging systems.

V. Conclusion

This study introduced a virtual imaging–guided harmonization method that standardizes chest radiographs by learning a transformation from non-harmonized to raw-like images. Trained on simulated XCAT data, the proposed GAN achieved substantial improvements in both generic and chest-specific image quality metrics and consistently outperformed ComBat and conventional denoising in reducing variability. When applied to clinical datasets, the harmonizer produced more compact and anatomically consistent latent feature spaces and improved diagnostic accuracy, particularly for findings driven by global contrast and structural cues. These outcomes indicate that the method effectively mitigates system-induced variability, improving both quantitative reproducibility and downstream AI robustness. The approach highlights the utility of virtual imaging trials for training deep harmonization models without large labeled datasets and offers a scalable path toward standardized radiographic appearance. Future extensions will expand the virtual cohort and integrate domain adaptation with clinical data to further enhance cross-population generalization and task-specific performance.

Fig. 1.

The generator network architecture comprised of two consecutive U-net networks: ReversalUNet and RestoralUNet.

Fig. 2.

Illustration of the utilized images during training. Non-harmonized images (red), ground truth (green), and harmonized images (blue).

TABLE I.

T-tests, effect sizes, and Bonferroni-corrected p-values between non-harmonized and harmonized CXRs for GIQMs.

	Software	T-stat	Effect Size	p-value
SSIM	A	9.709	2.171	3 × 10−14
	B	11.413	2.552	–
NRMSE	A	9.394	2.101	2 × 10−17
	B	8.706	1.947	–
PSNR	A	11.388	2.547	1 × 10−13
	B	12.961	2.898	–