Clinical feasibility and image quality assessment of deep learning-based virtual contrast enhancement from photon-counting computed tomography

Abstract

Objectives:

The purpose of this study was to evaluate image quality and diagnostic performance of deep learning-based virtual contrast-enhanced (VCE) images generated from photon-counting CT (PCCT) datasets.

Material and Methods:

Forty consecutive patients who underwent contrast-enhanced PCCT were retrospectively included and divided into training (n = 20) and testing (n = 20) cohorts. A supervised U-Net-based convolutional neural network was trained to generate VCE images from virtual non-contrast inputs, using true contrast-enhanced (TC) images as the reference. Quantitative similarity was assessed using the structural similarity index measure (SSIM), mean absolute error (MAE), mean squared error (MSE), and peak signal-to-noise ratio (PSNR). Qualitative evaluation was performed independently by two radiologists using a 5-point Likert scale across thoracic and abdominal structures, with assessment of exact, ±1-point, and diagnostic agreement.

Results:

Quantitative analysis demonstrated high similarity between VCE and TC images, with very high SSIM and PSNR values and low MAE, while MSE showed a right-skewed distribution. Qualitatively, VCE images were consistently rated lower than TC images. However, central thoracic vascular structures, cardiac chambers, and mediastinal lymph nodes were frequently rated as diagnostic (Likert ≥3) with high diagnostic agreement. In contrast, abdominal solid organs and vessels were often judged non-diagnostic, mainly due to heterogeneous enhancement and apparent contrast defects.

Conclusion:

The synthesis of VCE images from PCCT datasets is feasible and shows excellent quantitative similarity to TC images, with promising results for thoracic applications but limited reliability for abdominal evaluation at this stage.

INTRODUCTION

Contrast-enhanced computed tomography (CT) is essential to the routine radiological workflow, enabling improved visualization of vascular structures, solid organs, inflammatory processes, and malignant lesions. However, the use of iodinated contrast media is associated with potential adverse events, including hypersensitivity reactions and acute kidney injury, and may be contraindicated in specific patient populations.^[1,2] In addition, contrast-enhanced CT protocols may expose patients to higher radiation doses than their non-enhanced counterparts, which is of particular importance for patients undergoing repeated follow-up imaging.^[3,4]

Beyond its potential for spatial resolution, photon-counting CT (PCCT) substantially reduces electronic noise and provides spectral images of unprecedented quality.^[5,6] It also has the benefit of providing virtual non-contrast (VNC) reconstructions of higher quality, outperforming those obtained using energy-integrating detector CT (EIDCT).^[7]

Artificial intelligence has advanced substantially in recent decades, with deep learning (DL) emerging as a promising approach for medical image transformation and synthesis. Encoder–decoder architectures with skip connections, such as U-Net, have shown potential to perform image-to-image transformations while preserving fine details, making them useful for denoising, super-resolution and cross-modality image synthesis.^[8,9] More recently, DL models have been evaluated using clinical images to generate synthetic contrast-enhanced CT. These models have shown promise in thoracic imaging where they may improve the visualization of mediastinal lymph nodes and even raise suspicion of pulmonary emboli on unenhanced CT, and in abdominal imaging where they could improve diagnostic confidence and patient management.^[10-12]

However, most of these preliminary results were obtained using EIDCT data, which significantly differ from PCCTs in terms of noise, contrast, and image texture.^[13] To the best of our knowledge, no study to date has investigated the generation of virtual contrast-enhanced (VCE) images from PCCT datasets. The present work is therefore among the first to evaluate DL-based VCE in this setting.

Accordingly, this study aimed to evaluate the image quality and diagnostic performance of VCE CT images created from VNC PCCTs.

MATERIAL AND METHODS

Patients

We included 40 consecutive adult patients who underwent contrast-enhanced PCCTs with complete chest coverage (e.g., pulmonary CT angiography and thoracoabdominal CTs) between November 2024 and May 2025 and who had provided general consent for the use of their data in research. The study was approved by the Local Institutional Ethics Committee, and all procedures were conducted in accordance with the Declaration of Helsinki.

This sample was randomly divided into two sets: 20 patients were assigned to the training set, and 20 patients were assigned to the independent testing set.

Image acquisition

All examinations were performed on a PCCT system (NAEOTOM Alpha Pro, Siemens Healthineers, Erlangen, Germany). Although all scans were contrast-enhanced, acquisition parameters and contrast injection protocols varied according to the clinical indication.

Image processing

VCE images were generated from VNC PCCT datasets using a U-Net-based convolutional neural network. U-Net is a widely used encoder–decoder architecture for medical image analysis that incorporates skip connections to preserve spatial resolution and anatomical detail during image-to-image transformations.^[8] The UNet architecture used in this work consisted of four levels, each containing two convolutional layers followed by batch normalization and rectified linear unit activation. An input convolution projected the images to 64 features, which doubled at each level to a total of 1024. To limit the computational overhead, the feature maps were expanded using bilinear upsampling in the decoder part. The output convolution, which collapsed the feature map into a single-channel image, was implemented without an activation function to support a regression task rather than classification. The loss function comprised equally weighted L1 and perceptual losses.^[14] The perceptual loss was computed using features extracted from the 12^th layer of a VGG-19 network pretrained on ImageNet.^[15,16] The losses were optimized using the Adam optimizer^[17] for 200 epochs, with an initial learning rate of 0.01 that was reduced by a factor of 10 at epochs 30 and 60.

Consistency between VNC and TC images was ensured by the preprocessing steps. The pixel matrix was extracted from the DICOM file, and the pixel intensities were converted to Hounsfield units (HU) using the Rescale Intercept and Rescale Slope Modality Look-up table parameters (0028|1052 and 0028|1053, respectively). Pixel intensities were then clipped to the range (−1000, 2000) HU and linearly scaled as float32 to the (−1,1) interval. Finally, the images were downscaled to 256 × 256 pixels using bilinear interpolation. The images’ pixel intensities generated by the model were then mapped to (−1000, 2000) HU using the inverse linear transformation.

Supervised training of the model was then performed slice-by-slice using a VNC image as input and the corresponding true contrast-enhanced (TC) image as reference, allowing the model to learn voxel-wise contrast-enhancement patterns. In this setup, the model received no contextual information from surrounding slices.

Once the training was completed, the model was applied to the pre-processed VNC images of the test set, also on a slice-by-slice basis, to generate the VCE images of the test set in a DICOM format with the same slice thickness and reformatting as the original series, analyzable by the radiologists in a standard radiologic workflow.

An example of the VCE images compared with VNC and TC images is shown in Figure 1.

Figure 1:

Example of the image generation workflow in a 60-year-old female patient who underwent a thoraco-abdominal computed tomography with split-bolus contrast injection. (a-c) Axial images are shown at the level of the cervicothoracic junction, (d-f) pulmonary artery trunk, (g-i) and liver. All images are displayed using a soft-tissue window. The left column (a, d and g) shows virtual non-contrast images, the middle column (b, e and h) shows virtual contrast-enhanced images, and the right column (c, f and i) shows true contrast-enhanced images.

Image evaluation

Quantitative analysis

A quantitative evaluation of VCE images was performed by comparison with the corresponding TC images as the reference standard.

Image similarity was assessed using the structural similarity index measure (SSIM), which evaluates perceived image quality by integrating luminance, contrast, and structural information. In addition, voxel-wise intensity differences between VCE and TC images were quantified using mean absolute error (MAE), mean squared error (MSE), and peak signal-to-noise ratio (PSNR).

All quantitative metrics were computed on VCE/TC image pairs using whole-image comparisons.

Qualitative analysis

Two radiologists (R1, J.C., with 10 years of experience in thoracic imaging, and R2, D.L., with 1 year of experience) independently reviewed the TC and VCE images, with a 3-week washout period between readings to minimize recall bias. Radiologists could use all standard image-review tools available in routine clinical workflow. They were not blinded to the type of the images (TC or VCE), as the distinction is easily made by a radiologist.

For each examination, radiologists evaluated the following anatomical structures:

• Pulmonary vasculature, including proximal, segmental, and subsegmental pulmonary arteries, as well as pulmonary veins

• Cardiac chambers

• Intrathoracic systemic arteries

• Extrathoracic systemic arteries and veins

• Mediastinal lymph nodes

• Abdominal solid organs, including the liver, spleen, kidneys, and pancreas

• Abdominal vasculature, including the abdominal aorta, abdominal arteries, and abdominal veins

• Digestive tract

• Inguinal arteries and veins.

For each structure, radiologists assigned a qualitative Likert score to both TC and VCE images using the following criteria:

1. No or minimal contrast effect and/or severe image artifacts (e.g., pronounced heterogeneity, inappropriate contrast enhancement in normally non-enhancing structures), non-diagnostic

2. Low contrast and/or major artifacts; diagnostic confidence is very low

3. Suboptimal contrast and/or mild artifacts; diagnostic confidence is reduced

4. Good contrast enhancement with minor imperfections or slightly suboptimal attenuation; diagnostic confidence is high

5. Excellent contrast enhancement, close to the best enhancement possible for the analyzed structure.

Likert scores of 1–2 were considered non-diagnostic, while scores of ≥3 were considered diagnostic.

Statistics

Quantitative image similarity metrics (SSIM, PSNR, MAE, and MSE) were summarized using descriptive statistics.

Qualitative Likert scores were analyzed for each anatomical structure and for each reader separately. Agreement between TC and VCE images was assessed by calculating the proportion of identical Likert scores and the proportion of scores differing by no more than one point.

Diagnostic agreement was defined as agreement in diagnostic status between images judged diagnostic (Likert 3–4–5), and non-diagnostic (Likert 1–2).

Cohen’s kappa was not used, as the analysis focused on technique-to-technique agreement for ordinal ratings, and since skewed score distributions (in this case, large proportion of high Likert scores for the TC images) can yield misleadingly low kappa values despite high observed agreement.^[18]

Differences between TC and VCE scores were tested using the Wilcoxon signed-rank test. Interobserver variability between the two radiologists was evaluated using the same agreement measures.

p < 0.05 was considered statistically significant.

RESULTS

Examination types, radiation dose, and protocols

Patient demographics, CT acquisition parameters, radiation dose metrics, and contrast media characteristics for the training and testing cohorts are summarized in Table 1. The two cohorts were comparable in terms of age, sex distribution, and body habitus. Radiation dose levels were comparable across cohorts.

Table 1:

Patient characteristics, radiation dose metrics, and contrast administration parameters in the training and testing cohorts.

Variable	Training cohort (n=20)	Testing cohort (n=20)
Age (years)	63.5 (55.25–68.25)	62.5 (55.25–69)
Sex (male/female)	10 (50%)/10 (50%)	12 (60%)/8 (40%)
Weight (kg)	68.5 (59.25–78.75)	78 (59.25–85)
Height (m)	1.68 (1.58–1.73)	1.70 (1.66–1.74)
BMI (kg/m2)	24.84 (20.70–26.06)	26.99 (21.67– 28.73)
DLP (mGy•cm)	281 (180.25–452.25)	262.5 (159–444.5)
CTDIvol (mGy)	5.59 (5.19–6.87)	5.845 (4.755–7.08)
Injected contrast volume (mL)	100 (88.75–100)	100 (100–110)

Data are presented as median (Q1–Q3) (interquartile range) or number (%), unless otherwise specified. BMI: Body mass index, CTDI: Computed tomography dose index, DLP: Dose length product

Table 2 describes the different CT protocols used. CT examinations included a mix of chest and thoracoabdominal protocols, reflecting routine clinical practice. Arterial, venous, and split-bolus phase acquisitions were all represented.

Table 2:

Distribution of CT acquisition protocols in the training and testing cohorts.

CT acquisition protocol	Training cohort, n (%)	Testing cohort, n (%)
Thoracoabdominal split-bolus	8 (40)	10 (50)
Thoracoabdominal portal venous phase	3 (15)	1 (5)
Chest portal venous phase	4 (20)	3 (15)
Chest split-bolus	1 (5)	5 (25)
Chest hemoptysis protocol	4 (20)	1 (5)

CT: Computed tomography

Contrast material volumes were comparable between the training and testing cohorts. The type of contrast agent used was also similar across groups: 19 of 20 patients in each cohort received iohexol (Accupaque 350). In each cohort, the single remaining patient received a different iodinated contrast agent due to documented hypersensitivity to iohexol: iodixanol (Visipaque 270) in the training cohort and iomeprol (Iomeron 400) in the testing cohort.

Comparison between TC and VCE

Quantitative analysis

Table 3 presents the quantitative comparison between the proposed VCE model and the reference TC images.

Table 3:

Quantitative comparison between the proposed VCE model and true contrast images.

Metric	Mean±SD	Median (Q1–Q3)	Minimum	Maximum
MAE	28.11±3.16	28.41 (25.61–30.53)	20.49	35.74
MSE	919.6±1,470.3	457.0 (331.1–757.9)	117.6	9,598.7
PSNR (dB)	53.06±3.35	53.74 (51.74–55.20)	40.66	63.74
SSIM	0.99921±0.00271	0.99991 (0.99978–0.99994)	0.93504	0.99998

Metrics are computed on a per-instance basis and summarized across all test instances. Q1 and Q3 denote the first and third quartiles, respectively. MAE: Mean absolute error, MSE: Mean squared error, PSNR: Peak signal-to-noise ratio, SSIM: Structural similarity index measure, VCE: Virtual contrast enhancement, SD: Standard deviation

Structural similarity was very high with an SSIM of 0.99, indicating excellent preservation of anatomical structures. As for the contrast-related metrics, PSNR and MAE indicated good contrast fidelity. MSE, however, showed a right-skewed distribution, indicating a few contrast difference outliers.

Qualitative analysis

Overall, VCE images were consistently rated inferior to TC images across both readers and most anatomical structures [Tables 4-6], as reflected by systematically lower median Likert scores and low exact agreement.

Table 4:

Intra-observer agreement between true contrast-enhanced (TC) and virtual contrast-enhanced (VCE) images for Reader 1.

Anatomical structure	TC Likert	VCE Likert	Exact agm (%)	Agm±1 pt (%)	diagnostic agm (%)	p-value
Pulmonary arteries – proximal	4.0 (4.0–4.0)	4.0 (4.0–4.0)	55.0	95.0	100.0	0.275
Pulmonary arteries – segmental	4.0 (4.0–4.0)	4.0 (4.0–4.0)	55.0	95.0	100.0	0.275
Pulmonary arteries – subsegmental	4.0 (3.0–4.0)	3.0 (2.8–4.0)	60.0	95.0	85.0	0.042
Intrathoracic systemic arteries	4.0 (4.0–5.0)	4.0 (3.0–4.0)	40.0	95.0	100.0	0.023
Pulmonary veins	4.0 (4.0–4.2)	4.0 (3.8–4.0)	45.0	95.0	100.0	0.120
Cardiac cavities	4.0 (4.0–4.0)	4.0 (3.0–4.0)	50.0	90.0	95.0	0.015
Extrathoracic arteries	4.0 (4.0–5.0)	2.0 (2.0–2.0)	0.0	25.0	5.0	<0.001
Extrathoracic veins	2.0 (2.0–3.0)	2.0 (2.0–2.0)	50.0	90.0	70.0	0.145
Mediastinal lymph nodes	4.0 (3.0–4.0)	4.0 (3.0–4.0)	70.0	85.0	90.0	0.202
Liver	4.0 (4.0–4.0)	3.0 (3.0–3.0)	25.0	75.0	80.0	<0.001
Spleen	4.0 (4.0–4.0)	3.0 (2.0–3.0)	31.6	52.6	68.4	0.002
Kidney	5.0 (4.0–5.0)	2.0 (2.0–3.0)	5.6	22.2	38.9	<0.001
Pancreas	4.0 (4.0–4.2)	3.0 (3.0–4.0)	30.0	65.0	80.0	<0.001
Abdominal aorta	4.0 (4.0–5.0)	2.0 (2.0–3.0)	10.5	26.3	36.8	<0.001
Abdominal arteries	4.0 (4.0–5.0)	2.0 (2.0–2.0)	6.7	6.7	6.7	<0.001
Abdominal veins	3.0 (2.0–4.0)	2.0 (2.0–2.0)	35.0	65.0	40.0	0.001
Digestive tract	4.0 (4.0–4.0)	3.0 (2.8–4.0)	40.0	70.0	75.0	0.002
Inguinal arteries	4.0 (4.0–4.0)	2.0 (2.0–3.0)	0.0	36.4	36.4	0.003
Inguinal veins	2.0 (2.0–3.0)	1.0 (1.0–2.0)	36.4	81.8	63.6	0.018

Agm: Agreement, Agm±1: Agreement at more or less 1 Likert point. Diagnostic agm: Agreement between diagnostic (Likert 3–5) and non-diagnostic (Likert 1–2) status. Likert scores are reported as median (Q1–Q3); all agreement values are expressed as percentages. p-values are derived from Wilcoxon signed-rank tests comparing TC and VCE Likert scores; p < 0.05 was considered statistically significant.

Table 6:

Inter-observer agreement between the two radiologists for TC and VCE images.

Anatomical structure	Exact agm TC (%)	Agm±1 pt TC (%)	Diagnostic agm TC (%)	Exact agm VCE (%)	Agm±1 pt VCE (%)	Diagnostic agm VCE (%)
Pulmonary arteries – proximal	30.0	85.0	100.0	45.0	100.0	100.0
Pulmonary arteries – segmental	25.0	90.0	100.0	25.0	90.0	85.0
Pulmonary arteries – subsegmental	45.0	95.0	80.0	30.0	95.0	80.0
Intrathoracic systemic arteries	20.0	85.0	100.0	65.0	95.0	100.0
Pulmonary veins	25.0	90.0	100.0	60.0	95.0	95.0
Cardiac cavities	20.0	95.0	100.0	60.0	95.0	95.0
Extrathoracic arteries	30.0	85.0	100.0	0.0	95.0	95.0
Extrathoracic veins	20.0	70.0	35.0	10.0	100.0	100.0
Mediastinal lymph nodes	10.0	70.0	100.0	55.0	95.0	90.0
Liver	5.0	90.0	100.0	40.0	100.0	70.0
Spleen	21.1	84.2	94.7	30.0	85.0	55.0
Kidney	68.4	94.7	94.7	0.0	61.1	61.1
Pancreas	25.0	95.0	95.0	15.0	70.0	45.0
Abdominal aorta	35.0	100.0	100.0	26.3	73.7	36.8
Abdominal arteries	35.0	100.0	100.0	60.0	93.3	86.7
Abdominal veins	10.0	40.0	70.0	35.0	90.0	90.0
Digestive tract	25.0	100.0	100.0	30.0	70.0	50.0
Inguinal arteries	9.1	90.9	90.9	18.2	100.0	81.8
Inguinal veins	9.1	54.5	63.6	54.5	100.0	100.0

Agm: Agreement. Agm±1 : Agreement at more or less 1 Likert point. Diagnostic agm: Agreement between diagnostic (Likert 3–5) and non-diagnostic (Likert 1–2) status. All agreement values are expressed as percentages. TC: True contrast-enhanced, VCE: Virtual contrast-enhanced

Table 5:

Intra-observer agreement between TC and VCE images for Reader 2.

Anatomical structure	TC Likert	VCE Likert	Exact agm (%)	Agm±1 pt (%)	Diagnostic agm (%)	p-value
Pulmonary arteries – proximal	5.0 (5.0–5.0)	3.0 (3.0–4.0)	15.0	65.0	100.0	<0.001
Pulmonary arteries – segmental	5.0 (3.8–5.0)	3.0 (3.0–3.0)	15.0	60.0	85.0	<0.001
Pulmonary arteries – subsegmental	3.0 (3.0–4.0)	3.0 (2.8–3.0)	30.0	80.0	75.0	0.002
Intrathoracic systemic arteries	5.0 (5.0–5.0)	4.0 (4.0–4.0)	35.0	95.0	100.0	<0.001
Pulmonary veins	5.0 (4.8–5.0)	3.0 (3.0–4.0)	15.0	60.0	95.0	<0.001
Cardiac cavities	5.0 (5.0–5.0)	4.0 (3.0–4.0)	0.0	60.0	90.0	<0.001
Extrathoracic arteries	5.0 (5.0–5.0)	1.0 (1.0–1.0)	0.0	0.0	0.0	<0.001
Extrathoracic veins	3.0 (3.0–4.0)	1.0 (1.0–1.0)	0.0	5.0	5.0	<0.001
Mediastinal lymph nodes	5.0 (5.0–5.0)	4.0 (4.0–4.0)	5.0	85.0	100.0	<0.001
Liver	5.0 (5.0–5.0)	2.5 (2.0–3.0)	0.0	0.0	50.0	<0.001
Spleen	5.0 (5.0–5.0)	2.0 (1.8–3.0)	5.0	5.0	40.0	<0.001
Kidney	5.0 (5.0–5.0)	1.0 (1.0–1.0)	10.0	10.0	10.0	<0.001
Pancreas	5.0 (5.0–5.0)	2.0 (1.8–3.0)	5.0	10.0	40.0	<0.001
Abdominal aorta	5.0 (5.0–5.0)	3.0 (2.0–4.0)	0.0	35.0	65.0	<0.001
Abdominal arteries	5.0 (5.0–5.0)	2.0 (1.0–2.0)	0.0	0.0	5.0	<0.001
Abdominal veins	5.0 (5.0–5.0)	1.0 (1.0–2.0)	0.0	5.0	10.0	<0.001
Digestive tract	5.0 (5.0–5.0)	3.0 (2.0–3.0)	0.0	15.0	65.0	<0.001
Inguinal arteries	3.0 (1.0–5.0)	1.0 (1.0–2.0)	16.7	16.7	33.3	0.005
Inguinal veins	1.0 (1.0–4.0)	1.0 (1.0–1.0)	33.3	33.3	33.3	0.010

Agm: Agreement. Agm±1: Agreement at more or less 1 Likert point. Diagnostic agm: Agreement between diagnostic (Likert 3–5) and non-diagnostic (Likert 1–2) status. Likert scores are reported as median (Q1–Q3); all agreement values are expressed as percentages. TC: True contrast-enhanced, VCE: Virtual contrast-enhanced. p-values are derived from Wilcoxon signed-rank tests comparing TC and VCE Likert scores; p < 0.05 was considered statistically significant.

Examples of VCE images demonstrating excellent agreement and more moderate agreement with the corresponding TC datasets are shown in Figures 2 and 3, respectively.

Figure 2:

Example of good qualitative performance in a 78-year-old male patient who underwent a chest computed tomography with split-bolus contrast injection. (a-c) Axial images are shown at the level of the pulmonary arteries, (d-f) lower cardiac region, and (g-i) liver. All images are displayed using a soft-tissue window. The left column (a, d and g) shows virtual non-contrast images, the middle column (b, e and h) shows virtual contrast-enhanced (VCE) images, and the right column (c, f and i) shows true contrast-enhanced (TC) images. Both observers rated the pulmonary arteries and veins as well as the cardiac chambers as diagnostic on both examinations (VCE: Likert scores 3–5; TC: 4–5). The liver was rated as diagnostic by reader 1 (3/5), but as slightly non-diagnostic by reader 2 (2/5). The kidneys were rated as non-diagnostic by both radiologists. Notably, subsegmental lower lobe pulmonary emboli were well depicted on both VCE and TC (images e and f, respectively).

Figure 3:

Example of poorer qualitative agreement in a 67-year-old male patient who underwent a thoracoabdominal computed tomography with split-bolus contrast injection. (a-c) Axial images are shown at the level of the cardiac chambers, (d-f) liver, and (g-i) inferior kidneys, displayed using a soft-tissue window. The left (a, d and g), middle (b, e and h), and right (c, f and i) columns show virtual non-contrast, virtual contrast-enhanced (VCE), and true contrast-enhanced (TC) images, respectively. Thoracic vascular structures were rated as diagnostic on both VCE and TC images, whereas most abdominal structures were rated below the diagnostic threshold by both readers (Likert scores 1–2). Notably, heterogeneous hepatic enhancement (e) and an apparent partial filling defect of the abdominal aorta (h) reduced the image quality ratings in this case.

Thoracic structures

For central thoracic vascular structures, VCE demonstrated the best performance and closest agreement with TC.

Pulmonary arteries (proximal, segmental, and subsegmental), pulmonary veins, cardiac chambers, intrathoracic systemic arteries, and mediastinal lymph nodes consistently achieved moderate (50–74%) to high (≥75%) agreement within one Likert point for both readers. Binary diagnostic agreement for these structures was high (≥75%) in most cases, indicating that VCE images were frequently considered clinically interpretable when TC images were rated diagnostic.

Despite this, exact Likert agreement for thoracic structures remained mostly low, particularly for Reader 2, which appears to reflect systematic differences in the perceived degree of contrast enhancement rather than an actual loss of structural information.

Abdominal organs and vessels

In contrast, abdominal structures exhibited low to very low agreement between TC and VCE across all agreement metrics. Solid organs, including the liver, spleen, pancreas, and kidneys, showed low exact agreement and predominantly low or very low agreement within ±1 Likert point, with diagnostic agreement frequently falling below 50%.

Similarly, abdominal and inguinal vessels demonstrated very low agreement across all metrics. Exact agreement was frequently below 25%, agreement within ±1 Likert point was generally low, and binary diagnostic agreement was poor.

Inter-reader variability

Although both readers showed the same overall pattern, Reader 2 (the less experienced reader) tended to assign lower ratings to VCE images, particularly for abdominal structures, resulting in lower agreement levels and reduced diagnostic concordance compared with Reader 1. This trend is also visible in the Wilcoxon tests, which show predominantly non-significant differences for reader 1 but more frequent significant differences for reader 2.

However, the agreement of both readers within 1 Likert point was high to very high, as was the diagnostic agreement.

DISCUSSION

In this proof-of-concept study, our main result is that VCE images can be generated from PCCT VNC datasets using a deep-learning approach based on U-Net, with preserved diagnostic interpretability for most intrathoracic vascular structures but substantially degraded performance for abdominal organs and vessels. To the best of our knowledge, this work is among the first to evaluate DL-based VCE specifically from PCCT data, whereas previous synthetic contrast studies have relied almost exclusively on EIDCT.

The quantitative analysis demonstrated a high similarity between VCE and TC images, with consistently elevated SSIM and PSNR values and relatively stable MAE across test instances. The high SSIM values indicate excellent preservation of global structural information and spatial relationships, suggesting that the VCE model successfully maintains anatomical integrity. We noticed a right-skewed distribution of the MSE, which might indicate some errors in the VCE process, which is in line with the qualitative findings discussed below. More fundamentally, the contrast between the excellent quantitative metrics and the more modest qualitative ratings reflects what these indices actually measure. SSIM and PSNR are computed over the whole image and are dominated by overall anatomical structure, which the model reproduces faithfully. The differences that determine diagnostic quality, however, are differences in enhancement amplitude: an error of a few tens of HU in the liver parenchyma or in a vessel lumen is decisive for the radiologist, yet represents a small deviation across the wide dynamic range over which these indices are computed. Heterogeneous hepatic enhancement can therefore lower the qualitative rating of a large organ without appreciably reducing global similarity. These indices are also insensitive to the clinical relevance of an error: The apparent partial filling defect of the abdominal aorta shown in Figure 3 involves very few voxels and has a negligible effect on whole-image similarity, yet could be mistaken for a thrombus and would substantially alter diagnostic confidence. High quantitative similarity should consequently not be interpreted as evidence of diagnostic equivalence, and qualitative assessment by radiologists remains indispensable.

Concerning the qualitative findings, as expected, VCE images were consistently rated lower than TC images by both readers, reflecting reduced contrast intensity and altered image texture and/or synthetic artifacts. Importantly, despite these systematically lower ratings, the opacification of central thoracic structures such as pulmonary arteries (proximal to subsegmental branches), pulmonary veins, cardiac chambers, intrathoracic systemic arteries, and mediastinal lymph nodes remained diagnostically interpretable in most cases. These results are concordant with previous thoracic synthetic contrast studies on energy-integrating CT images, which report that large vessels, cardiac chambers, and mediastinal lymph nodes can remain clinically interpretable even when enhancement magnitude is reduced or the image is degraded by minor artifacts related to image transformation.^[10,11]

In contrast, and representing the principal limitation of the technique at this stage, abdominal structures showed a marked degradation of their opacification quality in the VCE images. Solid organs such as the liver, spleen, pancreas, and kidneys, as well as abdominal and inguinal vessels, frequently fell below the diagnostic threshold (Likert <3/5). These lower ratings may be attributed to several factors: Marked heterogeneity of the enhancement of the solid organ parenchyma and apparent endoluminal contrast defects in the vessels, which could mimic thrombi or atheromatous plaques. In some cases, aberrant contrast enhancement was observed in structures that are not expected to enhance.

The most likely explanation for these issues relates to the intrinsic complexity of abdominal contrast enhancement and the limitations of the current modeling strategy, which lacks contextual information from surrounding slices. In the chest, enhancement of central vessels and mediastinal structures follows relatively constrained anatomical rules, resulting in a near-deterministic mapping from VNC to contrast-enhanced appearance whereas in the abdomen, three-dimensional morphology and information from neighboring slices probably play a more prominent role in characterizing each organ. The fact that we used heterogeneous injection protocols in the training set may also have played a role and may have confounded the model due to the various possible enhancement patterns of the same organ at different phases. In our view, these considerations also help explain why other groups have reported more favorable abdominal results using EIDCT images, as their studies either incorporated contextual information (typically by providing sets of three adjacent slices), relied on more homogeneous contrast injection phases, or explicitly provided injection-phase information to the model.^[12,19]

The PCCT technique itself may also play a role in these differences due to different resolution, contrast and noise power spectrum, and the fact that PCCT VNC images are different from dual-energy integrating CTs.^[5,6] At the present stage, PCCT-based VCE images did not demonstrate a clear added value compared with approaches based on EIDCT datasets. Nevertheless, this work remains preliminary, and PCCT may ultimately offer a particularly favorable framework for synthetic contrast: Its intrinsic spectral capabilities and the more realistic VNC images it can generate may, in theory, provide additional leverage for training and optimizing such models in future studies.

Reader experience slightly modulated VCE assessment. The less-experienced reader consistently rated VCE images more severely, especially in the abdomen, whereas the more-experienced reader appeared more tolerant of reduced enhancement and non-standard texture. Such inter-observer differences, potentially related to reader experience and/or individual subjectivity, have been described across multiple AI-assisted imaging studies.^[20,21] In the present study, we believe these findings reflect differences in reliance on expected image appearance, with greater experience partially mitigating these concerns.

Beyond these technical considerations, the clinical rationale for VCE deserves to be stated explicitly. If such an approach reached maturity, its most immediate value would concern situations in which iodinated contrast media are undesirable or contraindicated, namely, patients with impaired renal function, documented hypersensitivity to iodinated contrast media, or thyroid disease, who currently undergo unenhanced CT with a reduced diagnostic yield. Similar considerations apply to patients requiring repeated follow-up imaging, in whom cumulative contrast exposure is a concern, and to situations of limited contrast media availability.

This study has several limitations. First, the sample size was small, particularly for training, which limited the diversity of enhancement patterns available to the model. This constraint may have reduced the model’s ability to generalize across the full spectrum of contrast dynamics encountered in routine clinical practice, especially for anatomically and physiologically complex regions such as the abdomen. The study was also conducted at a single center, on a single PCCT system, and with a limited range of reconstruction settings so that the model was trained and tested on data of homogeneous technical origin. No external validation cohort was available, and the generalizability of our results to other scanners, vendors, reconstruction algorithms, and patient populations therefore remains unknown. Multicenter studies including larger and more heterogeneous cohorts, together with external validation, will be indispensable before any clinical use can be envisaged.

Second, as discussed above, we chose slice-by-slice training and generation without volumetric context and bilinear feature map expansion in the decoder part. While this approach enabled reasonable computational efficiency and practical feasibility within our workflow, it constrained the network’s ability to capture long-range spatial relationships and complex organ–vessel interactions. In addition, as also previously mentioned, the variety of injection protocols used might have produced a one-to-many mapping problem during image generation, which may explain some of the less satisfactory results in abdominal organs.

Regarding the discrepancy between quantitative and qualitative results, although such differences are not uncommon in artificial intelligence studies, we believe that performing the quantitative analysis on whole images rather than organ-specific regions of interest may have slightly biased the quantitative metrics, particularly by reducing their sensitivity to localized contrast errors.

Finally, the use of VNC for training and test sets in this study may limit direct clinical applicability. However, this strategy, similarly to what other authors proposed,^[12] effectively eliminates biases related to motion artifacts and spatial misalignment. The clinically relevant target nonetheless remains the generation of synthetic contrast from true unenhanced acquisitions, which will require dedicated training on paired unenhanced and contrast-enhanced examinations and prospective evaluation against clinical endpoints.

Despite these limitations, the proposed approach yielded robust performance for thoracic mediastinal and vascular structures using PCCT input images, highlighting the potential of synthetic contrast enhancement techniques. Future work may benefit from increased computational resources to incorporate contextual or fully volumetric information, as well as from validation in routine clinical settings.

CONCLUSION

The synthesis of VCE images from PCCT datasets is feasible and yields excellent quantitative similarity to TC images. Performance was promising for thoracic applications but remained insufficiently reliable for abdominal evaluation at this stage, underscoring the need for further technical improvements and greater computational resources.

Funding Statement

Financial support and sponsorship: Nil.

Ethical approval:

The research/study approved by the Institutional Review Board at CCER, number 2025-00836, dated June 17, 2025.

Declaration of patient consent:

The authors certify that they have obtained all appropriate patient consent forms. In the form, the patient has given consent for their images and other clinical information to be reported in the journal. The patient understand that the patient’s names and initials will not be published and due efforts will be made to conceal their identity, but anonymity cannot be guaranteed.

Conflicts of interest:

There are no conflicts of interest.

Use of artificial intelligence (AI)-assisted technology for manuscript preparation:

The author(s) confirms that they have used artificial intelligence (AI)-assisted technology for language checks.