A generative adversarial network model for improved three-dimensional mapping of pulmonary arteries and veins from non-contrast computed tomography in sublobar resection planning

Introduction

Lung cancer remains one of the most prevalent and deadliest malignancies in the world (1). For patients with early-stage lung cancer, sublobar resection (including segmentectomy and wedge resection) has emerged as an effective surgical option, offering oncological outcomes comparable to those of lobectomy while providing greater preservation of lung function (2,3). The increasing use of low-dose computed tomography (CT) screening has led to a rise in the detection of small, early-stage lung nodules and thus a higher number of patients who are suitable for sublobar resection (4).

Precise preoperative planning for sublobar resection is technically demanding and depends on the surgeon’s ability to clearly visualize the patient’s individual bronchovascular anatomy. Three-dimensional (3D) reconstruction from preoperative CT scans has become an invaluable tool for achieving this end, providing a patient-specific anatomical roadmap that allows surgeons to identify the target segmental arteries, veins, and bronchi (5). This detailed mapping is essential to ensuring adequate surgical margins and preventing inadvertent damage to adjacent vascular structures.

Chest non-contrast CT (NCCT) is often insufficient for reliable 3D vascular reconstruction. The similar attenuation of pulmonary arteries (PAs) and pulmonary veins (PVs) on NCCT makes them difficult to distinguish from each other and from adjacent soft tissues, often resulting in fragmented and ambiguous vascular models that are inadequate for informed surgical decision-making (6). In contrast, chest contrast-enhanced CT (CECT) is the established standard, as intravenous contrast provides clear delineation of the intrapulmonary vasculature. However, CECT is contraindicated in a significant subset of patients due to factors such as severe iodine contrast allergy, renal insufficiency, and hyperthyroidism (7,8). This constitutes a critical deficiency in clinical practice, and there is an urgent need to develop a method for obtaining CECT-like vascular information from standard NCCT scans.

Artificial intelligence, particularly deep learning based on generative adversarial networks (GANs), offers a promising solution. As a well-established generative model, GANs demonstrate superior performance in medical image translation tasks, demonstrating remarkable success in synthetic CECT (Syn-CECT) images for improved visualization of aortic arteries and mediastinal lymph nodes (9,10). Recent studies have further confirmed that GANs are capable of generating realistic and detail-rich images to meet the demands of medical image synthesis. Diffusion models, representing the new generation of generative models, have also shown potential in terms of image quality and mode coverage (11), but they require powerful computing capabilities, which limits their practical application in many clinical and research scenarios.

To our knowledge, there are no reports on the application of GAN models for Syn-CECT images from NCCT scans specifically for the purpose of improving 3D bronchovascular reconstruction in sublobar resection planning. The purpose of this study was thus to develop a chest CT enhancement imaging model based on a GAN model, with NCCT images being used to generate high-quality synthetic CECT (Syn-CECT) images, and to quantitatively evaluate its clinical application value for improving vascular segmentation in the context of sublobar resection. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0279/rc).

Methods

Patient datasets

This retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Institutional Ethics Board of Zhongshan Hospital of Xiamen University (approval No. XMZSYY-AFSC-12-03). The requirement for informed consent was waived due to the retrospective nature of the analysis. For the training and validation sets, we included patients who underwent sublobar resection surgery and chest spectral CT angiography on an IQon scanner (Philips, Amsterdam, the Netherlands) at Zhongshan Hospital of Xiamen University from January 2022 to December 2023. Patients were excluded if there was poor CT image quality with severe motion artifacts, metallic implants causing significant radiation artifacts, or poor contrast agent visualization. The collected data were randomly divided into training and validation sets at an 8:2 ratio. To further evaluate model performance and clinical utility, two test sets were consecutively collected between October 2024 and January 2025 based on the CT scanner type (Figure 1). Test set 1 included patient scanned with an Ingenuity 128-slice spiral CT scanner (Philips), while test set 2 included patients scanned with a Revolution CT scanner (GE HealthCare, Chicago, IL, USA). All scans were performed at Zhongshan Hospital of Xiamen University. All datasets (training, validation, and the two test sets) were mutually exclusive at the patient level to ensure an unbiased evaluation of model generalization. Each patient’s clinical examination protocol included both a NCCT and a CECT phase, providing paired image sets. For spatial alignment, these paired NCCT and CECT images were then co-registered via itk-elastix (12).

Figure 1 Flowchart for the construction of the training and validation sets and the two test sets. 3D, three-dimensional; CECT, contrast-enhanced computed tomography; NCCT, non-contrast computed tomography; SBI, spectral base image; VNC, virtual non-contrast.

Image preprocessing

All axial CT images were downloaded in Digital Imaging and Communications in Medicine format from the picture archiving and communication system after de-identification. All CT images underwent linear interpolation resampling to achieve an image spacing of (1 mm × 1 mm × 1 mm), followed by central cropping to 256×256 pixels, and adjacent slices were stacked along the z-axis to create a 2.5-dimensional (2.5D) format, which was saved in Neuroimaging Informatics Technology Initiative (NIfTI) format. The 2.5D data included three cross-sectional layers (13). The original CT image range was from −1,024 to 3,071 Hounsfield units (HU). Since our study focused primarily on vascular compartment soft tissues, the original CT values were first clipped to −1,000 to 1,000 HU and then normalized to a range of −1 to 1.

Rationale for the training strategy: the use of paired spectral CT data

A key aspect of our training strategy was the use of spectral detector CT data. This technology allows for the material decomposition of a single post-contrast acquisition to generate a virtual non-contrast (VNC) image. The critical advantage is that the resulting VNC and CECT image pairs are perfectly co-registered at the pixel level, as they originate from the same scan and thus share the exact same anatomical position and motion profile. This ideal data pairing is crucial for training a supervised image-to-image translation model such as our GAN model, as it eliminates registration error as a confounding factor and allows the model to learn the direct mapping from nonenhanced to enhanced tissue characteristics. The model was subsequently tested on conventional NCCT and CECT scans from different scanners to validate its applicability in real-world clinical scenarios in which scans are acquired at different times.

Image annotation

The ground truth segmentations of the PA and PV on the CECT images were required for training the downstream segmentation model. These annotations were performed manually with 3D Slicer software. Initial annotations were completed by two radiology residents with 3 years of experience in thoracic imaging. All segmentations were subsequently reviewed and corrected by an attending cardiothoracic radiologist with over 10 years of experience to ensure accuracy. Following this, a no-new-Net (nnU-Net) segmentation model was trained using these labels (14). In the final stage, this trained segmentation model was deployed on the two separate test sets for automatic segmentation, with CECT serving as the gold standard, and this was followed by manual correction to eliminate errors to ensure precise segmentation results for the final evaluation.

GAN model

We directly used the Pix2pixGAN framework for our deep learning model, which consists of a generator and a discriminator network. The source code for training and inference is publicly available on GitHub (https://github.com/picklesdaddy/Enhance-GAN). The generator includes an attention UNet architecture. The integrated attention mechanism enables the network to dynamically focus on critical regions and fine-grained features. This design allows the generator to produce finer and more realistic CT details, thus improving the overall quality of Syn-CECT images, while also boosting training efficiency by suppressing irrelevant information during training (15). In terms of implementation, the decoder uses nearest-neighbor interpolation for upsampling to mitigate checkerboard artifacts, skip connections that link encoder block i with decoder block n−i are processed via rectified linear unit activation, and the final output layer employs a Tanh activation function. The discriminator is a PatchGAN network whose convolutional structure mirrors the encoder of the generator, and it distinguishes real from synthesized images by classifying 32×32×1 image patches.

GAN training

The GAN was developed in Python v.3.11 (Python Software Foundation, Wilmington, DE, USA) with PyTorch v.2.5.1. Training was conducted on a Linux workstation (Ubuntu 20.04) equipped with an RTX 3090 GPU (Nvidia, Santa Clara, CA, USA) with 24 GB of VRAM and 128 GB of RAM, requiring approximately 3 days to complete. The model was optimized via the AdamW algorithm. The learning rates for the generator and discriminator were set to 0.0005 and 0.0001, respectively, with β values of 0.5 and 0.999. The CosineAnnealingLR scheduler in PyTorch was applied to anneal the learning rates, with a T_max of 50 and eta_min values of 0.0001 for the generator and 0.00001 for the discriminator being set to promote stable convergence. Each epoch involved training on 40 randomly sampled 2.5D image groups over approximately 200 iterations. Our network was trained with a hinge adversarial loss and L1 distance loss (λ=10), and spectral normalization was applied to the discriminator’s convolutional layers to satisfy the Lipschitz constraint. The discriminator loss ( L_D) and generator loss (L_G) can be expressed as follows:

LD=E(x,y)[max(0,1−D(x,y))]+Ex[max(0,1+D(x,G(x)))][1]

LG=−Ex[D(x,G(x))]+λE(x,y)[‖y−G(x)‖1][2]

where x is the VNC, y is the true CECT, and the hyperparameter λ was empirically set to 10. To evaluate the performance of the generator, we objectively measured the quality of the synthesized images and assessed their suitability for a downstream 3D reconstruction task. The workflow of our approach is illustrated in Figure 2.

Figure 2 The workflow of the proposed approach. 2.5D, 2.5-dimensional; CECT, contrast-enhanced computed tomography; CT, computed tomography; NCCT, non-contrast computed tomography; nnU-Net, no-new-Net; Syn-CECT, synthetic CECT; VNC, virtual non-contrast.

Image registration

For the external test sets, sequentially acquired NCCT and CECT scans were aligned to compensate for varying breath-hold depths by a multistage 3D registration pipeline on itk-elastix. The pipeline sequentially applied the following: (I) a rigid transform (Euler transform and mutual information metric) for gross alignment; (II) an affine transform for global scaling/shearing; and (III) a nonrigid B-spline transform (normalized cross-correlation metric, a final grid spacing of 7 mm, and 5 multiresolution levels) to correct local lung parenchyma and vascular deformations.

Image quantitative evaluation

A quantitative evaluation was conducted on the two test sets through use of a suite of objective image quality metrics: the peak signal-to-noise ratio (PSNR) and the multiscale structural similarity index measurement (MS-SSIM). For these metrics, higher values indicate a greater similarity to the ground-truth images. Specifically, PSNR reflects the pixel-wise intensity differences, whereas MS-SSIM reflects the preservation of structural information. For a comprehensive comparison, these metrics were computed volumetrically across the entire 3D volume for both the Syn-CECT images and NCCT images, each relative to their corresponding ground truth.

Quantitative evaluation of preoperative 3D reconstruction for lung nodules

For the quantitative evaluation of preoperative 3D reconstruction accuracy, we used the nnU-Net model to automatically segment the PA and PV from NCCT, Syn-CECT, and CECT images. The CECT segmentations served as the ground truth. To specifically evaluate vessel completeness, structural integrity, boundary accuracy, and the local vascular performance around pulmonary nodules, we calculated three key metrics: the local Dice similarity coefficient (LocalDice), the fragmentation index (FragIndex), and the 95^th percentile Hausdorff distance (HD95). These metrics were computed for NCCT and Syn-CECT results against the CECT reference. All calculations were performed within a spherical region with a diameter of 50 mm centered on the pulmonary nodule. The LocalDice indicates the similarity between the segmented vessels and the ground truth and is calculated as follows:

LocalDice=2×TP/(FP+2×TP+FN)[3]

where TP is the number of true positives—the voxels correctly identified as vessels; FP is the number of false positives—the number of non-vessel voxels incorrectly identified as vessels; and FN is the number of false negatives—the number of vessel voxels missed by the segmentation. This metric directly reflects the accuracy of vascular segmentation in the area most relevant to surgical decision-making. The FragIndex measures the degree of fragmentation per unit of volume and is calculated as follows:

FragIndex=(N−1)/V[4]

where N is the number of disconnected vascular components, and V is the total vascular volume (cm³) in the target region. This metric specifically quantifies the vascular structural continuity around the nodule.

The HD95 evaluates the maximum surface boundary deviation between the segmented vessels and the ground truth, providing a rigorous measure of boundary accuracy that is critical for surgical planning.

A superior reconstruction is therefore characterized by higher LocalDice (indicating more consistent vascular segmentation with the ground truth), a lower FragIndex (indicating a more continuous, less fragmented vascular structure), and a lower HD95 (indicating a closer alignment of vessel boundaries with the ground truth and improved topological consistency).

Subjective evaluation of image quality

Two radiologists, each with 8 years of experience in thoracic imaging, performed subjective scoring on test sets 1 and 2, focusing on the ability to distinguish the PAs and PVs surrounding the pulmonary nodules on NCCT, Syn-CECT, and CECT images. In case of interobserver disagreement, a senior radiologist with 10 years of experience made the final decision for subsequent analysis. All radiologists, who were blinded to both the image acquisition method and the nodule pathology, performed the evaluation. Scoring was performed with a 5-point Likert scale according to the following scheme: 1, the PAs and PVs surrounding the pulmonary nodule are unidentifiable, with completely blurred vascular boundaries that cannot be distinguished from surrounding tissues; 2, the PAs and PVs surrounding the pulmonary nodule are difficult to distinguish, with blurred vascular boundaries that can only be barely identified after careful inspection; 3, the PAs and PVs surrounding the pulmonary nodule are partially distinguishable, with unclear vascular boundaries that can be discerned but with insufficient detail; 4, the PAs and PVs surrounding the pulmonary nodule are relatively clearly visualized, with distinct vascular boundaries that are easily identifiable and whose course can be traced; 5, the PAs and PVs surrounding the pulmonary nodule are clearly identifiable, with sharp vascular boundaries, smooth course, and clear contrast with surrounding tissues.

Statistical analysis

Statistical analyses were performed with R version 4.1.0 (https://www.r-project.org; The R Foundation for Statistical Computing, Vienna, Austria). The normality of distribution and homogeneity of variance for all quantitative metrics were assessed via the Shapiro-Wilk test and the Levene test, respectively. Data with a normal distribution and homogeneous variance are expressed as the mean ± standard deviation, while nonnormally distributed data were expressed as the median and interquartile range. For quantitative comparisons between NCCT and Syn-CECT, including LocalDice, FragIndex, and HD95, either the paired-samples t-test or the Wilcoxon signed-rank test was used as appropriate according to the data distribution. For subjective vascular identification score, comparisons between NCCT, Syn-CECT, and CECT in test sets 1 and 2 were performed with the Friedman test, followed by post hoc pairwise comparisons with the Bonferroni correction. Two-group comparisons were conducted via the Wilcoxon signed-rank test. A P value <0.05 was considered statistically significant.

Results

Patient characteristics

The patient characteristics and CT acquisition parameters are summarized in Table 1. A total of 212 patients were included in the study (119 men and 93 women; mean age 56.1±13.4 years; range, 22–96 years). This cohort comprised a training and validation set of 130 patients, which provided 34,115 image pairs (68,230 total images); test set 1 of 42 patients, with 10,734 image pairs (21,468 total images); and test set 2 of 40 patients, with 10,369 image pairs (20,738 total images). There were no statistically significant differences in clinical characteristics between the training and validation sets and test sets (P>0.05).

Quantitative image evaluation

Figure 3 presents a representative case from test set 2. As shown in Table 2, the Syn-CECT architectures (UNet, attention UNet, and SwinUNETR) consistently yielded higher MS-SSIM scores than did NCCT across both test sets (test set 1: 0.95–0.96 vs. 0.94; test set 2: 0.94–0.95 vs. 0.91). Conversely, the PSNR values for the Syn-CECT models were comparable to or marginally lower than those of NCCT (test set 1: 26.86–27.12 vs. 28.14 dB; test set 2: 26.36–26.66 vs. 26.73 dB). Notably, the performance metrics of the three Syn-CECT models were highly similar between the architectures. The 3D attention UNet model had lower PSNR and MS-SSIM values compared with UNet, attention UNet, SwinUNETR, and NCCT in both test set 1 (PSNR: 27.29 dB; MS-SSIM: 0.85) and test set 2 (PSNR: 26.44 dB; MS-SSIM: 0.84).

Figure 3 Images of a 51-year-old woman from test set 2 with a 1.2-cm invasive adenocarcinoma of the superior segment of right lower lobe. (A,D) NCCT; (B,E) Syn-CECT; (C,F) and CECT. All images are displayed with a lung window (window width =1,600 HU; window level =−550 HU). CECT, contrast-enhanced computed tomography; HU, Hounsfield units; NCCT, non-contrast computed tomography; Syn-CECT, synthetic CECT.

Evaluation of preoperative 3D CT visualization and reconstruction for pulmonary nodules

The nnU-Net model was employed to automatically segment the PA and PV from NCCT, Syn-CECT, and CECT images. The segmentation results from the CECT images were used as the reference ground truth. In test set 1 and test set 2, for the segmentation of PAs and PVs, Syn-CECT significantly outperformed NCCT. In both test sets, Syn-CECT, compared to NCCT, achieved higher LocalDice (0.89–0.90 vs. 0.85–0.89), lower FragIndex (3.82–5.92 vs. 4.54–8.68), and lower HD95 values (0.08 vs. 0.09–0.12), with all differences being statistically significant (P<0.01). The quantitative results are presented in Figure 4 and Table 3.

Figure 4 Comparison of NCCT and Syn-CECT on PA and PV segmentation according to the LocalDice, FragIndex, and HD95. (A) Results for test set 1. (B) Results for test set 2. CECT, contrast-enhanced computed tomography; FragIndex, fragmentation index; HD95, 95^th percentile Hausdorff distance; LocalDice, local Dice similarity coefficient; NCCT, non-contrast computed tomography; PA, pulmonary artery; PV, pulmonary vein; Syn-CECT, synthetic CECT.

Subjective evaluation of image quality

In test set 1, the median subjective vessel visibility scores for NCCT, Syn-CECT, and CECT were 3, 4, and 5, respectively, while in test set 2, the median scores were 3, 4, and 5, respectively. For both test sets, the overall differences between the three image types were statistically significant (P<0.01). Post hoc pairwise comparisons revealed that both Syn-CECT and CECT achieved significantly higher subjective scores than did NCCT (P<0.01) in both cohorts. All the statistical results are presented in Figure 5 and Table 4.

Figure 5 Vascular visualization performance of the three imaging modalities across both datasets. A 5-point scale was used for subjective scoring, with higher scores indicating better delineation of vascular structures on the images. In both test set 1 and test set 2, NCCT yielded relatively low vascular identification scores, with a median of 3. In comparison, Syn-CECT showed a significant improvement over NCCT, with a median score of 4. As the reference standard, CECT achieved the best performance in both test sets, with a median score of 5. Results of the Friedman nonparametric test included highly statistically significant differences in scores across the three imaging modalities. CECT, contrast-enhanced computed tomography; NCCT, non-contrast computed tomography; Syn-CECT, synthetic CECT.

Discussion

In this study, we developed and validated a GAN-based deep learning model to synthesize high-fidelity CECT images from routine NCCT scans. After first confirming that the Syn-CECT images had higher fidelity to true CECT than did the original NCCTs, we demonstrated a more clinically significant finding: when applied to the task of preoperative 3D reconstruction, use of Syn-CECT enabled significantly more accurate and complete segmentation of the pulmonary vasculature compared to use of the original NCCTs. This suggests that our approach can generate more precise 3D models of perinodular vasculature. Figure 6 shows a patient with a 1.2-cm invasive adenocarcinoma in the superior segment of the right lower lobe, accompanied by an anatomical variant of the right lower lobe artery—an accessory posterior basal segmental artery (indicated by the yellow arrow) arising between the dorsal segment artery and the basal trunk. On NCCT, this accessory posterior basal segmental artery appeared discontinuous because this part of the PA was locally adjacent to the PV, and some of the PA was missegmented as a PV. Meanwhile, Syn-CECT showed markedly improved vascular continuity, with visualization comparable to that of the reference-standard CECT. Misidentification of this arterial variant and unnecessary ligation of the accessory artery may lead to excessive pulmonary resection and unnecessary loss of functional lung parenchyma.

Figure 6 3D reconstruction of the pulmonary arterial tree (red) and pulmonary venous tree (translucent blue) in the same patient presented in Figure 3, derived from three imaging modalities: NCCT (left), Syn-CECT (middle), and true CECT (right). A pulmonary nodule (green) served as an anatomical landmark. The patient had a right lower lobe arterial variant, an accessory vessel known as the stellate segment artery, which is indicated by the yellow arrow and arises between the dorsal segment artery and the basal trunk. On NCCT, this stellate segment artery appeared discontinuous. In contrast, Syn-CECT demonstrated improved continuity of the vessel, with visualization comparable to that seen on the reference-standard CECT. 3D, three-dimensional; CECT, contrast-enhanced computed tomography; NCCT, non-contrast computed tomography; Syn-CECT, synthetic CECT.

We employed a 2.5D data-processing strategy in our model. This framework represents the 3D volume as a sequence of 2D slices while integrating spatial information from adjacent slices, thereby preserving interslice continuity. Although we also implemented and evaluated a full 3D processing architecture (3D attention UNet) with all hyperparameters matched except for batch size, our experimental results demonstrated that the 2.5D strategy achieved superior overall performance for Syn-CECT synthesis. Specifically, the 3D attention UNet yielded significantly lower PSNR and SSIM values, indicating inferior image quality. Compared with conventional pure 2D methods, the 2.5D strategy better captures 3D anatomical context and vascular topological continuity along the z-axis. Meanwhile, compared with the full 3D models, the 2.5D strategy substantially reduces computational memory consumption, allowing for a larger batch size during training; this not only enhances training stability but also contributes to a higher quality of generated Syn-CECT samples (16-19).

Traditional evaluation of pulmonary vascular segmentation mostly relies on global Dice or vascular abundance metrics (20,21). However, these global measures often fail to capture local performance in clinically critical regions or adequately reflect structural integrity. To address this, this study introduced LocalDice, FragIndex, and HD95 to better align the evaluation framework with the specific requirements of sublobar resection. These metrics facilitate optimizing “perinodular vascular assessment”, bridging the gap between computational metrics and detailed anatomical analysis. In two external validation sets, Syn-CECT demonstrated significant improvements over NCCT: the LocalDice increased by 0.01–0.04, the FragIndex decreased by 0.42–2.76, and the HD95 decreased by 0.01–0.02 mm (all P values <0.01). Although the absolute magnitude of these improvements appears slight, they indicate that Syn-CECT provides sharper vessel boundaries and reduced fragmentation. This improved delineation facilitates more reliable perinodular vascular visualization, offering a clearer anatomical reference for surgical planning compared to standard NCCT.

A cornerstone of our study’s success lies in the innovative training strategy that employs paired VNC and CECT images from a spectral CT scanner. Critically, VNC images closely approximate true NCCT, and studies have demonstrated excellent agreement in attenuation values and lesion conspicuity, thereby enabling their use as a valid surrogate for NCCT in clinical workflows (22,23). This approach provided perfectly co-registered ground truth data, effectively eliminating anatomical misalignment due to patient motion or breathing as a confounding variable (24). This allowed the GAN to learn the direct and intricate mapping of tissue enhancement patterns with high fidelity. The model’s robust generalization to conventional, non-spectral NCCT scans from two major vendors (Philips and GE HealthCare) attests to the efficacy of this strategy. Specifically, this suggests that the model learned the fundamental radiological principles of contrast enhancement rather than merely overfitting to the characteristics of a single scanner, which strongly supports its broad clinical applicability.

We should emphasize that our deep learning approach for generating Syn-CECT is not intended to replace CECT but rather to enhance the differentiation capability between PAs and PVs from standard NCCT, enabling the construction of higher-quality 3D vascular models for preoperative planning.

Limitations

Several limitations of this study should be acknowledged. First, although the training data from spectral CT provided perfectly co-registered VNC-CECT pairs at the pixel level, our external test sets consisted of sequentially acquired standard NCCT and CECT scans. Despite instructions for breath-hold at end-inspiration, physiological variations in breath-hold depth inevitably introduced local respiratory misalignments between the two phases, which might have partially affected the comparison results. Second, our study focused on a specific clinical application—pulmonary vascular segmentation for sublobar resection planning—and the model’s performance on other thoracic pathologies, such as mediastinal masses or interstitial lung disease, has not been tested. Finally, the single-center nature of our training data, although mitigated by testing on external datasets from two different scanner vendors (Philips and GE HealthCare), warrants further validation in larger, multicenter cohorts to ensure broad generalizability.

Conclusions

Our study developed a GAN-based model capable of Syn-CECT images from NCCT images. Compared with raw NCCT, the Syn-CECT images in this study provided moderate improvements in the accuracy and continuity of 3D pulmonary vascular reconstruction. This technique preliminarily demonstrates potential as an adjunctive tool for preoperative planning, and may provide additional vascular assessment information, particularly for patients with contraindications to iodinated contrast media. Further rigorous clinical validation is required to clarify the clinical value of this method and facilitate its practical application.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0279/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0279/dss

Funding: This study was funded by the Medical and Health Guidance Project Foundation of Xiamen City (No. 01105827).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0279/coif). H.Z. reports that this study was funded by the Medical and Health Guidance Project Foundation of Xiamen City (No. 01105827), and no other financial or non-financial interests exist. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by institutional ethics board of Zhongshan Hospital of Xiamen University (No. XMZSYY-AFSC-12-03) and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Saji H, Okada M, Tsuboi M, Nakajima R, Suzuki K, Aokage K, et al. Segmentectomy versus lobectomy in small-sized peripheral non-small-cell lung cancer (JCOG0802/WJOG4607L): a multicentre, open-label, phase 3, randomised, controlled, non-inferiority trial. Lancet 2022;399:1607-17. [Crossref] [PubMed]
3. Mamede I, Ribeiro L, Stecca C, Escalante-Romero L, Cypel M. Survival and pulmonary function in stage IA non-small cell lung cancer after sublobar resection versus lobectomy: An updated meta-analysis. J Surg Oncol 2024;130:523-32. [Crossref] [PubMed]
4. Bhamani A, Creamer A, Verghese P, Prendecki R, Horst C, Tisi S, et al. Low-dose CT for lung cancer screening in a high-risk population (SUMMIT): a prospective, longitudinal cohort study. Lancet Oncol 2025;26:609-19. [Crossref] [PubMed]
5. Cannone G, Verzeletti V, Busetto A, Lione L, Bonis A, Nicotra S, Rebusso A, Mammana M, Schiavon M, Dell'Amore A, Rea F. Three-Dimensional Imaging-Guided Lung Anatomic Segmentectomy: A Single-Center Preliminary Experiment. Medicina (Kaunas) 2023;59:2079. [Crossref] [PubMed]
6. Wu X, Liu SJ, Li FW, Xin X, Chen YT. Application Research of Chest Plain Scan and Enhanced CT Assisted 3D Reconstruction in Early NSCLC Translational Medicine Journal 2024;13:1807-10.
7. Li Y, Wang J. Contrast-induced acute kidney injury: a review of definition, pathogenesis, risk factors, prevention and treatment. BMC Nephrol 2024;25:140. [Crossref] [PubMed]
8. Berglund F, Eilertz E, Nimmersjö F, Wolf A, Nordlander C, Palm F, Parenmark F, Westerbergh J, Liss P, Frithiof R. Acute and long-term renal effects after iodine contrast media-enhanced computerised tomography in the critically ill-a retrospective bi-centre cohort study. Eur Radiol 2024;34:1736-45. [Crossref] [PubMed]
9. Lyu J, Fu Y, Yang M, Xiong Y, Duan Q, Duan C, Wang X, Xing X, Zhang D, Lin J, Luo C, Ma X, Bian X, Hu J, Li C, Huang J, Zhang W, Zhang Y, Su S, Lou X. Generative Adversarial Network-based Noncontrast CT Angiography for Aorta and Carotid Arteries. Radiology 2023;309:e230681. [Crossref] [PubMed]
10. Choi JW, Cho YJ, Ha JY, Lee SB, Lee S, Choi YH, Cheon JE, Kim WS. Generating synthetic contrast enhancement from non-contrast chest computed tomography using a generative adversarial network. Sci Rep 2021;11:20403. [Crossref] [PubMed]
11. Kazerouni A, Aghdam EK, Heidari M, Azad R, Fayyaz M, Hacihaliloglu I, Merhof D. Diffusion models in medical imaging: A comprehensive survey. Med Image Anal 2023;88:102846. [Crossref] [PubMed]
12. Ntatsis K, Dekker N, Van Der Valk V, Birdsong T, Zukić D, Klein S, Staring M, McCormick M. itk-elastix: Medical image registration in Python. In: Proceedings of the 22nd Python in Science Conference. 2023:101-5.
13. Zhang Y, Liao Q, Ding L, Zhang J. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph 2022;99:102088. [Crossref] [PubMed]
14. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021;18:203-11. [Crossref] [PubMed]
15. Zhao J, Hou X, Pan M, Zhang H. Attention-based generative adversarial network in medical imaging: A narrative review. Comput Biol Med 2022;149:105948. [Crossref] [PubMed]
16. Kumar A, Jiang H, Imran M, Valdes C, Leon G, Kang D, Nataraj P, Zhou Y, Weiss MD, Shao W. A flexible 2.5D medical image segmentation approach with in-slice and cross-slice attention. Comput Biol Med 2024;182:109173. [Crossref] [PubMed]
17. Bi R, Guo L, Yang B, Wang J, Shi C. 2.5 D cascaded context-based network for liver and tumor segmentation from CT images. Electronic Research Archive 2023;31:4324.
18. Avesta A, Hossain S, Lin M, Aboian M, Krumholz HM, Aneja S. Comparing 3D, 2.5D, and 2D Approaches to Brain Image Auto-Segmentation. Bioengineering (Basel) 2023;10:181. [Crossref] [PubMed]
19. Wang J, Qin Z, Zhang Y, Hu VT, Ommer B, Briq R, Kesselheim S. Scaling Image Tokenizers with Grouped Spherical Quantization. arXiv:2412.02632 [Preprint]. 2024. Available online: https://arxiv.org/abs/2412.02632
20. Chu Y, Luo G, Zhou L, Cao S, Ma G, Meng X, Zhou J, Yang C, Xie D, Mu D, Henao R, Setti G, Xiao X, Wu L, Qiu Z, Gao X. Deep learning-driven pulmonary artery and vein segmentation reveals demography-associated vasculature anatomical differences. Nat Commun 2025;16:2262. [Crossref] [PubMed]
21. Xu S, Zhang Z, Zhou Q, Shao W, Tan W. A Pulmonary Vascular Extraction Algorithm from Chest CT/CTA Images. J Healthc Eng 2021;2021:5763177. [Crossref] [PubMed]
22. Steinhardt M, Marka AW, Ziegelmayer S, Makowski M, Braren R, Graf M, Gawlitza J. Comparison of Virtual Non-Contrast and True Non-Contrast CT Images Obtained by Dual-Layer Spectral CT in COPD Patients. Bioengineering (Basel) 2024;11:301. [Crossref] [PubMed]
23. Mingkwansook V, Puwametwongsa K, Watcharakorn A, Dechasasawat T. Comparative study of true and virtual non-contrast imaging generated from dual-layer spectral CT in patients with upper aerodigestive tract cancer. Pol J Radiol 2022;87:e678-87. [Crossref] [PubMed]
24. Yang Y, Liu J, Zhan G, Chen Q, Wang F, Li Y, Kumar Jain R, Lin L, Hu H, Chen YW OA-GAN. organ-aware generative adversarial network for synthesizing contrast-enhanced medical images. Biomed Phys Eng Express 2024; [Crossref]