Diffusion Model for Generating Synthetic Contrast Enhanced CT

from Non-Enhanced Heart Axial CT Images

Victor Rogerio Sousa Ferreira

, Anselmo Cardoso de Paiva

, Aristofanes Correa Silva

João Dallyson Sousa de Almeida

, Geraldo Braz Junior

and Francesco Renna

Universidade Federal do Maranhão, Av. dos Portugueses, 1966, Bacanga,

São Luís/MA, Núcleo de computação Aplicada, UFMA, São Luís, Brazil

INESC TEC, Faculdade de Ciências da Universidade do Porto, Porto, Portugal

Keywords: Diffusion Models, Image Translation, Adversarial Networks.

Abstract: This work proposes the use of a deep learning-based adversarial diffusion model to address the translation of

contrast-enhanced from non-contrast-enhanced computed tomography (CT) images of the heart. The study

overcomes challenges in medical image translation by combining concepts from generative adversarial

networks (GANs) and diffusion models. Results were evaluated using the Peak signal to noise ratio (PSNR)

and structural index similarity (SSIM) to demonstrate the model's effectiveness in generating contrast images

while preserving quality and visual similarity. Despite successes, Root Mean Square Error (RMSE) analysis

indicates persistent challenges, highlighting the need for continuous improvements. The intersection of GANs

and diffusion models promises future advancements, significantly contributing to clinical practice. The table

compares CyTran, CycleGAN, and Pix2Pix networks with the proposed model, indicating directions for

improvement.

1 INTRODUCTION

Non-communicable diseases (NCDs) are medical

conditions that cannot be spread directly from one

person to another and are often caused by a

confluence of behavioural, physiological,

environmental, and genetic variables.

In (WHO, 2023), it is stated that NCDs are the world's

greatest cause of mortality, accounting for 41 million

deaths per year (74% of all deaths worldwide). The

bulk of NCD-related fatalities (17.9 million/year) are

attributable to cardiovascular diseases, which also

have a significant role in premature death and

disability globally (Dondi, 2021).

A timely and efficient way to improve population

health overall is through diagnostic imaging. By early

discovery, they can be utilised as a preventive

approach to lessen cardiovascular issues.

A common imaging modality for diagnosing

cardiovascular disorders is computed tomography

(CT) (Corballis, 2023; Counselor, 2023)—an

imaging modality with growing diagnostic utility.

Cardiac CT plays a crucial role in diagnosing and

managing heart diseases. It is possible to obtain

detailed three-dimensional images of the heart

through cardiac CT, allowing for precise evaluation

of cardiac anatomy, function, and circulation. This

makes cardiac CT a valuable tool for diagnosing

various heart conditions, including coronary artery

disease, cardiomyopathies, and congenital heart

defects

Cardiac CT is particularly useful in detecting

coronary artery disease (CAD), one of the leading

causes of morbidity and mortality worldwide.

Coronary artery disease (CAD) is a significant

cardiovascular disease defined by a narrowing or

blockage of the coronary arteries. With cardiac CT,

the existence and severity of CAD may be evaluated

non-invasively. Coronary calcium deposits, linked to

an elevated risk of coronary artery disease, are often

evaluated by non-contrast cardiac computed

tomography (NCCT). On the other hand, contrast-

enhanced cardiac computed tomography (CECT) is

used when the goal is to quantify cardiovascular

disease, evaluate blood flow dynamics, define the

composition of plaque, and offer quantitative

measurements of disease severity. This tomography

technique injects contrast materials into the body to

increase the contrast between certain organs, blood

Ferreira, V., Cardoso de Paiva, A., Silva, A., Sousa de Almeida, J., Braz Junior, G. and Renna, F.

Diffusion Model for Generating Synthetic Contrast Enhanced CT from Non-Enhanced Heart Axial CT Images.

DOI: 10.5220/0012724600003690

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 26th International Conference on Enter prise Information Systems (ICEIS 2024) - Volume 1, pages 857-864

ISBN: 978-989-758-692-7; ISSN: 2184-4992

857

arteries, or tissues and the surrounding structures on

CT images.

CECT improves patient outcomes by helping

physicians detect and track many elements of

cardiovascular disease. It does this by making

cardiovascular structures and abnormalities more

visible.

But in contrast to NECT, CECT is more costly,

involves more radiation exposure, and may have

unfavourable side effects, including headaches and

vomiting. Furthermore, anyone with allergies or renal

problems might be in danger when using CECT.

There has been great potential for using artificial

intelligence in cardiac CT to improve diagnosis and

prognosis. This exam has unique characteristics that

make it even more attractive for the application of

artificial intelligence, although the complexity of this

application is increasing.

Another application is in the assessment of

cardiac function from cardiac CT images. Algorithms

can automatically analyse cardiac volumes, ejection

fraction, and wall motion, providing precise

measurements that aid in evaluating heart function.

This automated analysis saves time and resources,

allowing physicians to focus on more complex

interpretations and personalised treatment plans for

each patient.

Among these challenges, we highlight the need

for quantitative evaluations, which generally involve

quantitative evaluations such as ventricular volume,

fraction of blood volume ejected out of the ventricle,

volumetric evaluation of heart muscle tissue, amount

of plaque present in the coronary arteries, area of

stenosis, among others. In addition, the images are

acquired with thinner slices, and the evaluation

targets are smaller (e.g., coronary arteries).

Given the potential risks associated with CECT,

generative AI-based techniques for cardiac imaging

can assist professionals in assessing coronary artery

disease without the drawbacks of CECT. More

precisely, we can create a CECT image that matches

the given NECT image using data-driven methods

without requiring contrast substance injection.

The challenge of medical image synthesis may be

approached through picture-to-image translation

(Parmar, 2023) or style transfer (Jing, 2020). This

topic poses extra complications in the context of

cardiac CT images. Since the same patient's NECT

and CECT images are frequently significantly out of

alignment, direct monitoring for NECT to CECT

mapping is rarely feasible.

In recent years, medical image translation has

emerged as a powerful solution to overcome these

challenges. This process involves synthesising

images of the target modality based on the guidance

of images acquired from the source modality.

However, the inherent nonlinear variations in tissue

signals between modalities make this problem

complex and ill-conditioned.

Learning-based methods, especially Generative

Adversarial Networks (GANs), have shown

remarkable success in image translation tasks. GANs

employ an adversarial mechanism in which a

discriminator guides a generator to perform a one-

time mapping to produce the target image. While

GANs exhibit exceptional realism in image synthesis,

they indirectly characterize the target modality

distribution, potentially introducing biases and

limiting the mapping process's reliability.

As an alternative approach, recent studies in

computer vision have explored diffusion models

based on explicit likelihood characterisation and a

gradual sampling process to enhance sample fidelity.

However, the potential of diffusion methods in

medical image translation remains largely

unexplored, partly due to computational challenges

and difficulties in the non-paired training of regular

diffusion models.

In this work, we propose a deep learning-enabled

image-to-image translation model that can map

contrast-free CT images of the heart to contrast-

enhanced ones. To achieve this, we implemented an

adversarial diffusion model, applying concepts from

GANs to generate high-quality images. This method

aims to provide an accurate model compared to other

approaches.

2 RELATED WORKS

In (Azarfar, 2023), authors present several papers

proposing deep learning architectures to reduce or

eliminate administered contrast media to acquire

clinically useful computer tomographies.

The introduction of GANs (Goodfellow, 2014)

presented an innovative approach to generative

models. GANs operate based on the principle of

rivalry between two networks - the generator and the

discriminator. The generator aims to produce

synthetic data indistinguishable from real data, while

the discriminator strives to differentiate between the

two. Through adversarial training, GANs achieve

Nash equilibrium, converging the generator's

distribution to the training data.

In image translation, especially in the analysis of

medical images, GANs are widely used for their

ability to automatically learn patterns in input data so

that the model can generate new examples (output)

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

858

that could exist in the original dataset. When

performing image generation, the simplest model

maps from source to destination through a trained

generator using adversarial loss (Goodfellow, 2014).

Consequently, the GAN-based translation approach

has been extensively adopted in various applications.

Conditional GANs excel in mapping a single

source to a destination, improving sensitivity to high-

frequency details in tissue structure compared to

traditional pixel-to-pixel losses. Integrating

adversarial loss terms has proven effective in

enhancing spatial accuracy and realism in target

images synthesised with GANs, surpassing

conventional convolutional models.

Other studies, specifically using GANs, whose

primary focus lies in the synthesis of contrast-

enhanced computed tomography (CECT) images

from non-contrast CT (NCT) scans, are (Chun, 2022)

and (Seo, 2021). They employ a two-stage framework

and sophisticated network architectures as generators,

such as DenseNet and SPADE (Park, 2019). They

successfully align NCT and CECT images,

surpassing previous methods in accuracy and

applicability.

Other research extends to artery-contrasted

computed tomography (ACT), which is crucial for

diagnosing conditions like aneurysms. To mitigate

the risks of contrast agents, they introduce an aorta-

aware deep learning approach that synthesises artery-

contrasted CT volumes directly from non-contrast CT

data (Hu, 2022). Utilising aGANs and innovative loss

functions, their model demonstrates remarkable

accuracy in estimating ACT slices, thus enhancing

diagnostic precision while minimising patient risk.

The Pix2Pix (Zhu, 2017) model is one of the

approaches designed for image-to-image translation

tasks. It consists of a generator to create synthetic

images and a discriminator to distinguish between

real and generated images. Training involves an

adversarial process, where the generator tries to

deceive the discriminator, and the discriminator seeks

to identify fake images.

Applications of the Pix2Pix architecture (Choi,

2021) utilise the fundamental structure of the original

pix2pix model to generate synthetic contrast

enhanced from non-contrast chest CT, with the

distinction that the 2D convolutional layers are

substituted by their 3D equivalents. This model

comprises a generator and discriminator networks

akin to a conventional GAN. The generator network

is a U-Net convolutional neural network encoder-

decoder with skip connections. The discriminator

network is a PatchGAN that classifies each pixel

patch as real or fake, and its convolutional module is

identical to the encoder block of the generator.

A dissertation (Domingues, 2022) compares the

performance of two GAN models, Pix2Pix-GAN and

Cycle-GAN, in generating contrast-enhanced images

from non-contrast CT scans. The study explores the

trade-offs of using 2D, 2.5D, and 3D inputs,

employing different types of generators and datasets.

Evaluation metrics include Structural Similarity

Index Measure (SSIM), Peak Signal-to-Noise Ratio

(PSNR), Mean-Square Error (MSE), and Dice metric

for high contrast region fidelity.

CyTran (Ristea, 2021) is a GAN-based model

designed for working with CT images. This

innovative approach focuses on bidirectional

translation of contrast and non-contrast computed

tomography (CT) scans, even when the images lack

direct pairing. CyTran aims to address the challenge

of generating contrast scans for patients who cannot

receive contrast and to enhance the alignment

between contrast and non-contrast CT scans. The

method employs a cycle-consistent architecture based

on generative adversarial transformers designed for

transferring CT scans across different contrast

phases. Inspired by the CycleGAN framework,

CyTran comprises two discriminators and two

generators, enabling training on unpaired images

through a multi-level cycle consistency loss. In

addition to ensuring image-level consistency, Cytran

utilises additional losses between intermediate

feature representations to enhance the model's

performance further. This comprehensive strategy

contributes to the model's effectiveness in translating

CT images bi-directionally, offering valuable

applications in medical imaging scenarios.

However, GANs present their challenges. Issues

such as lower reliability in mapping a single sample,

premature convergence of the discriminator, and poor

representational diversity leading to mode collapse

can compromise the quality and diversity of

generated samples. Despite these challenges, GANs

currently lead in image generation tasks, surpassing

other models based on metrics such as Inception

Score and Accuracy.

Lately, deep diffusion models have become an

alternative to GANs in generative modelling tasks in

computer vision (Yang, 2022). These models are

inspired by non-equilibrium thermodynamics,

defining a Markov chain of diffusion steps to add

random noise to the data slowly and then learning to

reverse the diffusion process to construct desired data

samples from the noise. Noise removal is conducted

by a neural network architecture trained to maximise

the correlation between adjacent pixels.

Diffusion Model for Generating Synthetic Contrast Enhanced CT from Non-Enhanced Heart Axial CT Images

859

This diffusion technique provides greater

reliability in network mapping and improves the

quality and diversity of generated samples. The two-

step structured diffusion model starts with direct

diffusion, where input data is gradually perturbed

over multiple steps by adding Gaussian noise. In the

reverse step, the model is trained to recover the

original data, reversing the diffusion process step by

step. This innovative method offers a robust and

effective approach to generating realistic data in

various computer vision contexts.

Moreover, diffusion models are easily adaptable,

able to use different architectures such as

Transformers (Peebles, 2023) and adversarial

networks (Wang, 2023), achieving results that

surpass the quality of previous diffusion models in

metrics like peak signal-to-noise ratio (PSNR), ratio

is used as a quality measurement between the original

and a compressed image, and structural similarity

index measure (SSIM) for measuring the similarity.

Table 1: Overview of the applied techniques cited in related

works.

Related Works Applied Techniques

(Ristea, 2021) CycleGan structure with

Pix2Pix + Transformers

(Seo, 2021) GAN

(

SPADE +DCGAN

)

(Chun, 2022) GAN

(FC-DenseNet + PatchGAN)

(

Choi, 2021

)

Pix2Pix

(Domingues, 2022) CycleGan and Pix2Pix with

SkipResidual Generato

(

Park, 2019

)

SPADE

3 METHODOLOGIES

This work proposes a methodology for synthesising

NECT to CECT images, utilising a GAN-based

approach with diffusion models. This methodology

consists of the following steps: data acquisition, data

pre-processing and proposed network architecture.

3.1 Data Acquisition

The dataset was obtained at the Orca Score in the

Grand Challenge platform (Wolterink,2022). Images

in this dataset were acquired on four different CT

scanners from four different vendors in four different

hospitals using standard parameters for calcium

scoring in cardiac CT. For each patient, both a non-

contrast-enhanced CT and a contrast-enhanced

computed tomography angiography (CTA) image are

provided. The training set consists of images of 32

patients. The test set consists of images of 40 patients.

From this dataset, 6209 images were extracted,

divided into 2812 for testing and 3397 for training and

validation. The entire set consists of images with and

without contrast from the same patients.

3.2 Data Pre-Processing

For this study, it was essential to conduct specific

preprocessing steps before utilising these CT images

to enhance the overall quality of the model. The

preprocessing involved segregating the slices of

contrast and non-contrast CT images of each patient's

heart, selecting those in the same position with a high

similarity index. This approach ensured that only the

most relevant and corresponding images were used to

refine the model's analysis.

To achieve this, the images were correlated using

the SSIM and Normalised Cross-Correlation (NCC)

to assess the structural similarity.

Images with higher similarity indices were

subsequently considered equivalent. Following this,

the best images from each patient, meaning those with

the same position and the highest similarity indices

correlating contrast and non-contrast, were separated

and allocated into training, testing, and validation

sets. The number of retained images was as follows.

Table 2: Orca Dataset Paired Filtration Summary.

Contrast Non-Contrast

Train 200 200

Test 100 100

Validation 50 50

3.3 Proposed Network Architecture

Based on the SynDiff network (Özbey, 2023), a

diffusion model was developed with a conditional

origin adversarial projector for fast and accurate

reverse diffusion sampling. Unlike conventional

models that use a relatively large number of steps, this

network employs fast-forward diffusion, adaptively

adjusting noise variance to balance efficiency and

precision in image generation.

The proposed network utilises a Cycle-GAN

architecture consisting of diffusive generators and a

non-diffusive discriminator (Figure 1). The diffuse

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

860

Figure 1: Architecture of proposed network.

generator translates images from the NECT to the

CECT domain and vice versa. Conversely,

discriminators aim to distinguish between real and

generated images. In the diffusive module, generators

employ a UNet backbone comprising six encoding

and decoding blocks (Ho, 2020). Each block includes

two residual subblocks followed by a convolutional

layer. During encoding, the convolutional layer

reduces the feature map resolution by half, while the

channel dimensionality is doubled every other block.

The convolutional layer doubles the resolution for

decoding, while the channel dimensionality is halved

every other block.

The proposed network utilises a Cycle-GAN

architecture consisting of diffusive generators and a

non-diffusive discriminator (Figure 1). The diffuse

generator translates images from the NECT to the

CECT domain and vice versa. Conversely,

discriminators aim to distinguish between real and

generated images. In the diffusive module, generators

employ a UNet backbone comprising six encoding

and decoding blocks (Ho, 2020). Each block includes

two residual subblocks followed by a convolutional

layer. During encoding, the convolutional layer

reduces the feature map resolution by half, while the

channel dimensionality is doubled every other block.

The convolutional layer doubles the resolution for

decoding, while the channel dimensionality is halved

every other block.

The discriminator model is designed as a

sequential neural network (Radford, 2015), tailored

for input images of size 256 by 256 pixels. It consists

of two convolutional layers with 64 and 128 filters,

each with a kernel size of (5, 5) and a stride of (2, 2)

for downsampling. Leaky ReLU activation functions

introduce non-linearity after each convolutional

layer. To prevent overfitting, dropout layers with a

dropout rate of 0.3 are incorporated after each Leaky

ReLU layer.

Given the larger input dimensions, the

architecture is adapted to handle the increased spatial

information. Following the convolutional layers, a

flattening layer transforms the 2D feature maps into a

1D vector. Finally, a dense layer with one neuron is

added, serving as the output layer for binary

classification (discriminating between real and

generated CT images).

During training, the proposed network enforces

cycle consistency, a crucial property that ensures the

translated images maintain semantic content and

realism. By incorporating an additional loss function

to quantify the disparity between the NECT image

generated by the second generator and the original

NECT image, as well as vice versa, the proposed

network promotes cycle consistency. This

regularisation technique guides the generator models

in the creation of CECT images. The generators aim

to minimise both the adversarial loss, which measures

their ability to generate realistic images, and the

cycle-consistency loss simultaneously. Meanwhile,

discriminators are trained to improve their ability to

distinguish between real and generated CT images,

thereby providing feedback to the generators.

The training objective of the proposed network

resembles the CycleGAN method, which utilises two

main loss functions: adversarial loss and cycle-

consistency loss. Adversarial loss incentivises the

generators to produce images indistinguishable from

real images, as perceived by the discriminators. On

the other hand, cycle-consistency loss enforces the

constraint that translating an image from one domain

to another and then back should result in a

reconstruction close to the original image

4 EXPERIMENT AND RESULTS

To evaluate the adequacy of the proposed

architecture, we conducted an experiment using the

Orca dataset and compared the results with other

papers that employ GAN approaches to generate

CECT images from NECT images..

The network hyperparameters were set as

follows: 100 epochs, the Adam optimizer with beta1

= 0.5 and beta2 = 0.9, a learning rate of 10^-4 for the

diffusion method and GAN, a batch size of 2, T =

1000, which represents the number of interactions in

the noising and denoising process, a step size of k =

250, and T/k = 4 diffusion steps. The weight loss in

diffusion and cycle models was set to λ1φ = 0.5

Diffusion Model for Generating Synthetic Contrast Enhanced CT from Non-Enhanced Heart Axial CT Images

861

The metrics were obtained through the

comparison between generated CECT images with

real ones. The metrics presented in Table 3 are based

on the averages of these results.

The results obtained, exemplified by Figure 2 and

Table 3, showcase the remarkable performance of the

proposed diffusion model. Notably, the

competitiveness of the PSNR and SSIM indicators in

generating contrast-enhanced heart images reflects

the model's significant ability to preserve both quality

and visual similarity.

Figure 2: Images with contrast generated by the network

from non-contrast images.

A noteworthy point is that although the images

have a high degree of visual similarity, MAE and

RMSE values are still much higher than expected. A

good example is the two images below, which exhibit

considerable visual resemblance but yield MAE and

RMSE values as high as 0.6 and 0.7, respectively.

Figure 3: Real Contrast Images (Left) and the Generated

One (Right).

However, in other images, the MAE and RMSE

values reached 0.11 and 0.15, respectively,

demonstrating that depending on the image, the

network can generate a more accurate version closer

to the real one.

Figure 4: A Feature Matching of Real Contrast Images

(Left) and the Generated One (Right).

However, upon analysing the Root Mean Square

Error (RMSE) values, it is observed that, despite the

visual resemblance of the generated images, the

model predictions deviate significantly from the

actual values, as seen in Table 3.

Upon closer examination of the images, a subtle

yet discernible variance in the absolute pixel values

between the original and the generated samples

becomes apparent. Indeed, a slight disparity exists

between the generated pixel values and the

corresponding ideal pixel values, as shown in Figure

2, with the generated pixel values exhibiting a

marginally higher magnitude. While these minor

discrepancies may seem inconspicuous individually,

their cumulative effect in the summation process

significantly contributes to the observed dissimilarity

reflected in the RMSE.

It is important to note that the interpretation of

RMSE depends on the specific domain of the problem

and the units of the variable being predicted. In some

cases, a high RMSE may be acceptable if it aligns

with the natural variations in the data or is justified by

the nature of the problem being addressed. These

indicators suggest the presence of substantial

variations that require a deeper understanding of the

generated images. Studying these discrepancies can

provide valuable insights further to enhance the

effectiveness of the image generation process.

Table 3: Comparison of results among CyTran, Cycle-

GAN, Pix2Pix-GAN networks, Cycle-GAN-2D, and

Cycle-GAN-2D with SkipResidual Generator against the

proposed model.

Model MAE RMSE SSIM PSRN

CyTran 0.061 0.144 0.745 29.66

cle-GAN 0.066 0.150 0.724 29.22

Pix2Pix-GAN 0.070 0.165 0.729 29.51

cle-GAN-2D 0,030 - 0,433 15,569

Pix2Pix-GAN-2D 0,025 - 0,492 16,375

Proposed

Diff-Model

0.061 0.200 0.701 32,85

5 CONCLUSIONS

This paper addresses the translation of contrast and

non-contrast cardiac computed tomography (CT)

images using a deep learning-based adversarial

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

862

diffusion model. By overcoming challenges

associated with medical image translation, we

explore an approach that combines concepts from

generative adversarial networks (GANs) and

diffusion models. The obtained results, evaluated

through metrics such as PSRN and SSIM, showcase

the remarkable capability of the model in generating

contrast-enhanced cardiac images while preserving

quality and visual similarity. However, the analysis of

RMSE indicates persistent challenges, suggesting the

presence of variations that require a deeper

understanding to enhance the consistency and fidelity

of the generated images.

In conclusion, the developed model delivers

notable results, but the study acknowledges the need

for continuous improvements to address variations in

the generated images. The intersection of GANs and

diffusion models proves promising, pointing towards

future research and developments in medical image

translation and significantly contributing to

advancing this crucial area in clinical practice.

ACKNOWLEDGEMENTS

The authors acknowledge the Coordenação de

Aperfeiçoamento de Pessoal de Nível Superior

(CAPES), Brazil - Finance Code 001, Conselho

Nacional de Desenvolvimento Científico e

Tecnológico (CNPq), Brazil, and Fundação de

Amparo à Pesquisa Desenvolvimento Científico e

Tecnológico do Maranhão (FAPEMA) (Brazil),

Empresa Brasileira de Serviços Hospitalares (Ebserh)

Brazil (Grant number 409593/2021-4), and the

Portuguese funding agency, FCT - Fundação para a

Ciência e a Tecnologia, within project

UIDB/50014/2020.DOI.10.54499/UIDB/50014/202

0 | https://doi.org/10.54499/uidb/50014/2020 for the

financial support.

REFERENCES

Azarfar, G., Ko, SB., Adams, S.J. et al. (2023) Applications

of deep learning to reduce the need for iodinated

contrast media for CT imaging: a systematic review. Int

J CARS 18, 1903–1914. https://doi.org/10.1007/s115

48-023-02862-w

Choi, J.W., Cho, Y.J., Ha, J.Y. et al. (2021) Generating

synthetic contrast enhancement from non-contrast chest

computed tomography using a generative adversarial

network. Sci Rep 11, 20403.

https://doi.org/10.1038/s41598-021-00058-3

Chun, J., Chang, J. S., Oh, C., Park, I., Choi, M. S., Hong,

C. S., ... & Kim, J. S. (2022). Synthetic contrast-

enhanced computed tomography generation using a

deep convolutional neural network for cardiac

substructure delineation in breast cancer radiation

therapy: a feasibility study. Radiation Oncology, 17(1),

1-9.

Corballis, N.; Tsampasian, V.; Merinopoulis, I.;

Gunawardena, T.; Bhalraam, U.; Eccleshall, S.; Dweck,

M.R.; Vassiliou, V. CT.(2023) angiography compared

to invasive angiography for stable coronary disease as

predictors of major adverse cardiovascular events—A

systematic review and meta-analysis. Heart Lung, 57,

207–213.

Counseller, Q. and Aboelkassem, Y. (2023) Recent

technologies in cardiac imaging, Frontiers in Medical

Technology, Vol. 4, DOI 10.3389/fmedt.2022.984492,

Croitoru, F. A., Hondru, V., Ionescu, R. T., & Shah, M.

(2023). Diffusion Models in Vision: A Survey, in IEEE

Transactions on Pattern Analysis and Machine

Intelligence, vol. 45, no. 9, pp. 10850-10869.

Dondi M, Paez D, Raggi P, Shaw LJ, Vannan

M.(2021).Integrated non-invasive cardiovascular

imaging: a guide for the practitioner. International

Atomic Energy Agency.

Domingues, R. A. D. (2022). Automatic contrast generation

from contrastless CTs, Master Thesis, Universidade do

Porto, FCUP - Faculdade de Ciências.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2020).

Generative adversarial networks. Communications of

the ACM, 63(11), 139-144.

H., & Kwak, S. (2021). Neural contrast enhancement of CT

image. In Proceedings of the IEEE/CVF Winter

Conference on Applications of Computer Vision (pp.

3973-3982).

Henry, J., Natalie, T., & Madsen, D. (2021). Pix2Pix GAN

for Image-to-Image Translation. Research Gate

Publication, 1-5.

Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion

probabilistic models. Advances in neural information

processing systems. Editors: H. Larochelle and M.

Ranzato and R. Hadsell and M.F. Balcan and H. Lin},

Vol. 33, pages 6840--685}, Curran Associates, Inc.},

https://proceedings.neurips.cc/paper_files/paper/2020

/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

Hu, T., Oda, M., Hayashi, Y., Lu, Z., Kumamaru, K. K.,

Akashi, T., ... & Mori, K. (2022). Aorta-aware GAN for

non-contrast to artery contrasted CT translation and its

application to abdominal aortic aneurysm detection.

International Journal of Computer Assisted Radiology

and Surgery, 1-9. https://doi.org/10.1007/s11548-021-

02492-0

Jing, Y., Yang, Y., Feng, Z., Ye, J., Yu, Y., & Song, M.

(2019). Neural style transfer: A review. IEEE

transactions on visualization and computer graphics,

26(11), 3365-3385.

Karras, T., Aittala, M., Aila, T., & Laine, S. (2022).

Elucidating the design space of diffusion-based

generative models. Advances in Neural Information

Processing Systems, 35, 26565-26577.

Diffusion Model for Generating Synthetic Contrast Enhanced CT from Non-Enhanced Heart Axial CT Images

863

Park, T., Liu, M. Y., Wang, T. C., & Zhu, J. Y. (2019).

Semantic image synthesis with spatially-adaptive

normalization. In Proceedings of the IEEE/CVF

conference on computer vision and pattern recognition

(pp. 2337-2346).

Parmar, G., et all. (2023). Zero-shot Image-to-Image

Translation. In ACM SIGGRAPH 2023 Conference

Proceedings (SIGGRAPH '23). Association for

Computing Machinery, New York, NY, USA, Article

11, 1–11. https://doi.org/10.1145/3588432.3591513

Peebles, W. and Xie, S. (2023) Scalable Diffusion Models

with Transformers," IEEE/CVF International

Conference on Computer Vision (ICCV), Paris, France,

2023, pp. 4172-4182, doi: 10.1109/ICCV51070.20

23.00387

Radford, A., Metz, L., & Chintala, S. (2015). Unsupervised

representation learning with deep convolutional

generative adversarial networks. arXiv preprint

arXiv:1511.06434.

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net:

Convolutional networks for biomedical image

segmentation. In Medical Image Computing and

Computer-Assisted Intervention–MICCAI 2015: 18th

International Conference, Munich, Germany, October

5-9, 2015, Proceedings, Part III 18 (pp. 234-241).

Springer International Publishing.

Ristea, N. C., Miron, A. I., Savencu, O., Georgescu, M. I.,

Verga, N., Khan, F. S., & Ionescu, R. T. (2021). Cytran:

Cycle-consistent transformers for non-contrast to

contrast ct translation. arXiv preprint arXiv:2110.064

00.

Seo, M. et al., "Neural Contrast Enhancement of CT

Image", 2021 IEEE Winter Conference on Applications

of Computer Vision (WACV), Waikoloa, HI, USA,

2021, pp. 3972-3981, doi: 10.1109/WACV48630.20

21.00402.

Özbey, M., Dalmaz, O., Dar, S. U., Bedel, H. A., Özturk,

Ş., Güngör, A., & Çukur, T. (2023). Unsupervised

medical image translation with adversarial diffusion

models. IEEE Transactions on Medical Imaging.

Wang, Z., Zheng, H., He, P., Chen, W., & Zhou, M. (2022).

Diffusion-gan: Training gans with diffusion. arXiv

preprint arXiv:2206.02262.

Wolterink, J., Vo, B.D., Leine, T. , Viergever, M. A.,

Išgum, I.(2022) Orca score. https://orcascore.grand-

challenge.org

WHO World Health Organization. Noncommunicable

diseases. (2023) Available online: https://www.who.

int/news-room/fact-sheets/detail/noncommunicable-

diseases.

Zhu, J. Y., Isola, P .Zhou, T., & Efros, A. A. (2017). Image-

to-image translation with conditional adversarial

networks. In Proceedings of the IEEE conference on

computer vision and pattern recognition (pp. 1125-

1134).

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

864