Joint Denoising and Motion-Correction for Low-Dose CT Myocardial Perfusion Imaging Using Deep Learning

Abstract

Computed Tomography (CT) is a widely used imaging modality that employs X-rays and computational reconstruction to visualize internal anatomy. Although higher radiation doses produce higher-quality images, they also increase long-term cancer risk, motivating the use of low-dose protocols. However, low-dose CT data inherently suffer from elevated Poisson–Gaussian noise, necessitating effective denoising strategies. In myocardial CT perfusion (CTP) imaging, this challenge is compounded by residual cardiac motion, which misaligns consecutive time points and impairs accurate estimation of perfusion maps for diagnosing coronary artery disease. Traditional approaches typically treat these two problems, noise and motion, separately, denoising the reconstructed images first or applying the registration first. Such serial pipelines often degrade clinically significant features; e.g., denoising may destroy structural details essential for registration, while motion correction can distort subtle intensity cues needed for noise modelling. To overcome these limitations, we propose a unified deep learning framework that performs noise suppression and motion correction jointly for low-dose myocardial CTP. The method integrates two complementary components through a parallel ensemble strategy: (i) a modified Fast and Flexible Denoising Network (FFDNet) that incorporates noise-level maps to mitigate blended noise effectively, and (ii) a CNN-based registration model, extended with Time Enhancement Curve (TEC) correction and 4D physiological consistency constraints to estimate temporally coherent and anatomically plausible motion fields. By combining their outputs without iterative dependencies, the proposed framework produces motion-corrected and denoised CTP sequences in a single unified processing step, thereby better preserving myocardial structure and perfusion dynamics than conventional serial pipelines. The model has been evaluated using both reference-based (MSE, PSNR, SSIM, PCC, Noise Variance, TRE) and no-reference (NIQE, FID, KID, AUC) image quality metrics, supplemented by expert human assessment. Results demonstrate that jointly learning noise characteristics and motion patterns enables restoration of low-dose CTP images while minimizing feature corruption, thereby advancing the clinical utility of low-dose myocardial CTP imaging.

1. Introduction

Myocardial CTP imaging is a powerful modality for evaluating coronary artery disease (CAD), providing both qualitative and quantitative information on myocardial blood flow (MBF) through dynamic contrast-enhanced acquisitions [1,2,3,4]. Unlike static CT, myocardial CTP captures a temporal sequence of volumetric images over multiple cardiac phases following contrast injection, enabling assessment of perfusion defects, hemodynamic significance, and ischemic burden. However, this dynamic acquisition process of myocardial CTP introduces two fundamental challenges that compromise image quality and quantitative reliability: (i) increased noise due to low-dose radiation protocols, and (ii) residual cardiac and respiratory motion causing frame-to-frame misalignment. Addressing both issues simultaneously is essential for preserving perfusion dynamics and maintaining image characteristics relevant to diagnostic interpretation.

1.1. Noise in Low-Dose Myocardial CTP

Because dynamic CT perfusion imaging requires repeated scanning across several cardiac cycles, radiation dose management is a primary clinical concern [5]. Reducing X-ray tube current or voltage is a widely adopted strategy to mitigate dose; however, this unavoidably increases Poisson–Gaussian noise in the reconstructed images [6]. Elevated noise levels degrade visual quality, obscure minor perfusion defects, and can bias quantitative metrics such as MBF and myocardial blood volume [7].

Traditional noise-reduction methods range from linear filtering [8], nonlinear diffusion [9,10], and compressed sensing [11] to vendor-specific model-based iterative reconstruction schemes such as ASIR and ADMIRE [12,13]. Although effective in reducing noise, these techniques often blur fine anatomical structures, introduce texture inconsistencies, or require proprietary reconstruction software.

Deep learning (DL)–based image denoising has emerged as a powerful alternative, offering improved noise suppression while preserving anatomical fidelity [14,15]. Architectures such as DnCNN [14], UNet [16], and GAN-based models [17] have demonstrated promising results in CT, MRI, and PET. However, existing approaches rarely address the unique challenges of myocardial CTP, including temporal consistency, dynamic contrast behaviour, and the need to preserve small myocardial structures [18].

1.2. Motion in Dynamic Myocardial CTP

In addition to noise, myocardial CTP is highly susceptible to motion artifacts arising from cardiac contraction, respiration, and involuntary patient movements [19,20]. While ECG gating and breath-hold instructions help mitigate some motion, they cannot fully eliminate frame-to-frame misalignment, particularly in patients with irregular rhythms or limited compliance. Even sub-voxel misalignments can distort time enhancement curves (TECs) and impair MBF estimation, potentially masking subtle perfusion defects in the subendocardium—the region most sensitive to ischemia [21].

Motion correction in cardiac imaging is typically achieved using deformable image registration, as rigid or affine transforms cannot capture the complex nonlinear myocardial motion. Classical registration methods, including optical flow, B-splines, and diffeomorphic frameworks, estimate voxel-wise deformation fields to align dynamic frames [22]. However, their direct application to myocardial CTP is challenged by rapid contrast changes (violating brightness constancy assumptions), confounding anatomy within the field of view, and a lack of ground-truth deformation fields for supervised training [23,24].

Recent advances in learning-based registration, including CNNs and transformer-based architectures [25,26], have shown promise by learning motion patterns directly from data. Although effective in lung, abdominal, and cardiac PET/SPECT applications [27,28], their validation in myocardial CTP remains limited. The dynamic contrast kinetics and rapid motion interplay in CTP demand modality-specific adaptations to preserve TEC integrity and MBF quantification accuracy.

1.3. Motivation for Noise and Motion Correction

Noise suppression and motion correction in myocardial CTP are tightly coupled. Performing denoising first may remove subtle structural cues essential for accurate motion estimation, while performing registration first may distort intensity patterns critical for noise modelling. Traditional serial pipelines, therefore, risk degrading perfusion-relevant features that subsequent stages rely on. We conducted a Serialized Versus Unified Correction ablation study to test whether a unified approach is indeed preferable to conventional serialized pipelines. We compared three strategies on compound simulations (

Table 1. Performance comparison for serialized corrections versus unified correction under noise and motion degradation ($\overline{x}$ represents mean, and $\sigma$ represents standard deviation).

Test Case	MSE						PSNR						SSIM
	Noise/ Motion Corr.		Motion/ Noise Corr.		Comp. Correction		Noise/ Motion Corr.		Motion/ Noise Corr.		Comp. Correction		Noise/ Motion Corr.		Motion/ Noise Corr.		Comp. Correction
	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$
STC1	8.66	3.41	8.73	4.14	6.79	3.35	38.76	1.84	38.72	1.87	40.30	2.05	0.90	0.00	0.90	0.00	0.91	0.00
STC2	7.30	3.14	7.41	3.82	5.62	3.25	39.50	2.03	39.43	2.06	41.22	2.19	0.90	0.00	0.90	0.00	0.91	0.00
STC3	8.56	3.87	8.70	4.64	6.78	4.16	38.80	2.05	38.74	2.08	40.43	2.23	0.90	0.00	0.90	0.00	0.91	0.00
STC4	8.22	3.23	8.32	3.91	6.41	3.28	38.98	1.94	38.93	1.97	40.58	2.11	0.90	0.00	0.90	0.00	0.91	0.00
STC5	6.88	2.04	6.97	2.49	5.31	2.05	39.75	1.26	39.70	1.28	41.13	1.41	0.89	0.00	0.89	0.00	0.90	0.00
STC6	8.64	2.47	8.77	2.98	6.78	2.66	38.77	1.22	38.70	1.25	40.07	1.40	0.88	0.00	0.88	0.00	0.89	0.00
STC7	7.12	2.18	7.23	2.67	5.39	2.36	39.61	1.39	39.54	1.41	41.11	1.51	0.89	0.00	0.89	0.00	0.90	0.00
STC8	9.10	2.76	9.24	3.34	7.07	2.90	38.54	1.30	38.48	1.33	39.90	1.42	0.88	0.00	0.88	0.00	0.89	0.00

):

• Noise correction followed by motion correction,
• Motion correction followed by noise correction,
• Unified compound correction (proposed).

Across all test cases, the unified approach achieves the lowest MSE, highest PSNR, and highest SSIM. Typical improvements over serialized pipelines are modest but consistent. For example, in STC1, MSE decreases from ∼8.7 (serialized) to $6.79$ (unified), PSNR increases from ∼38.7 dB to 40.3 dB, and SSIM from ∼0.90 to 0.91. Similar trends are observed for all STC cases. Between the two serialized variants, applying denoising before registration yields slightly better metrics than the reverse order, suggesting that reducing noise may facilitate more reliable motion estimation. However, both serialized strategies are consistently outperformed by the unified framework, supporting the central hypothesis of this work: addressing noise and motion within a joint, ensemble-based formulation preserves more task-relevant information than sequential preprocessing and leads to reconstructions that are closer to high-dose, motion-free acquisitions in both static and dynamic characteristics.

This observation motivates the development of a method that jointly addresses noise and motion, tailored to dynamic myocardial CTP. In this study, we present a unified noise and motion correction framework specifically optimized for myocardial CTP.

2. Related Work

Research on low-dose myocardial CTP intersects three major themes: (i) noise suppression for low-dose CT acquisitions, (ii) motion correction via deformable registration in dynamic cardiac imaging, and (iii) emerging frameworks that seek to address perfusion-specific objectives. Below, we review each of these areas.

2.1. Noise Suppression in Low-Dose CT and Myocardial CTP

Noise suppression in low-dose CT (LDCT) has been extensively investigated as a means to balance the radiation–image-quality trade-off. Iterative reconstruction (IR) emerged as a significant advance over conventional filtered back projection (FBP), repeatedly refining image estimates using projection data to reduce noise and improve contrast [29,30]. Over time, IR has evolved into statistical, adaptive statistical, model-based, and hybrid variants [31,32,33], enabling substantial dose reductions while maintaining diagnostic utility. However, IR methods are computationally expensive, often require proprietary access to raw projection (sinogram) data, and may exhibit texture changes or over-smoothing at ultra-low-dose levels, where photon statistics become highly non-Gaussian and structural detail is challenging to preserve [13,34,35]. Compressed sensing (CS) introduced sparsity priors to reconstruct high-quality images from undersampled data [11,36], but remains limited by long reconstruction times and sensitivity to parameter tuning, which have curtailed widespread clinical adoption [37].

More recently, deep learning–based denoising has reshaped the LDCT landscape, both in the reconstruction and image domains [6,30,37]. Reconstruction-domain methods operate directly on sinograms or intermediate representations, suppressing noise prior to image formation. Examples include sinogram interpolation and denoising frameworks such as SIPID [38], domain-progressive residual networks that jointly process sinogram, FBP, and image domains [39], and primal–dual U-Net architectures for sinogram upsampling in undersampled CT [40]. While these methods demonstrate strong technical performance, they are constrained by limited access to raw projection data and vendor-specific formats, which limit large-scale training and multi-center deployment.

Image-domain denoising has therefore become the most widely studied and implemented strategy, as reconstructed CT images are readily available in clinical practice. Early CNN-based approaches, such as Chen et al. [15], demonstrated significant gains in PSNR and SSIM over classical filters, but often oversmoothed fine anatomical structures. Subsequent work introduced deeper residual encoder–decoder networks [41], directional wavelet CNNs [42], and structurally sensitive loss functions combining SSIM, adversarial, and pixel-wise terms to preserve edges and textures better [43,44,45]. GAN-based methods further emphasized perceptual quality and realistic texture reproduction, albeit sometimes at the expense of traditional quantitative metrics [17,46]. Additional studies have evaluated CNN-based denoising for emphysema screening [47], quantitative CT imaging [48], and domain-specific LDCT protocols [8,9,29,31,32,49].

DnCNN [14] and its variant FFDNet [50] are among the most influential architectures in this space. DnCNN models residual noise, allowing the network to focus on learning noise patterns rather than clean intensities, while FFDNet introduces spatially varying noise-level maps for greater flexibility. These models have been adapted to LDCT and other modalities [47,48], and more recent work has extended them with transformer-based components [51,52]. However, standard formulations are primarily designed for additive Gaussian noise, whereas ultra-low-dose CT data exhibit a mixed Poisson–Gaussian distribution arising from photon statistics and electronic noise [53,54]. Self-supervised/unsupervised training strategy [55] has been explored to bridge this gap, but robust handling of realistic mixed noise remains an open challenge, particularly in dynamic myocardial CTP, where temporal consistency and perfusion patterns must also be preserved.

Although substantial progress has been made in denoising CT images from the thorax, abdomen, and other regions [6,32,33,56], relatively few studies have explicitly focused on myocardial CTP. Kadimesetty et al. [57] applied CNN-based denoising directly to CT perfusion maps rather than raw images, and Ramon et al. [58] used a 3D autoencoder to predict high-dose myocardial perfusion images from low-dose inputs. However, most work still treats myocardial CTP as a downstream application of generic LDCT denoising rather than a primary target with its own temporal and perfusion-specific constraints.

Noise degradation in low-dose myocardial CTP was previously addressed using an extended FFDNet-based denoising framework [59]. In that work, the FFDNet architecture was adapted to explicitly model the mixed Poisson–Gaussian noise characteristics of CTP data by incorporating a Generalized Anscombe Transform to stabilize signal-dependent noise into an approximately Gaussian distribution. The study demonstrated robust noise suppression in low-dose myocardial CTP, motivating the use of this denoising model as a core component of the present unified motion- and noise-correction framework.

2.2. Motion Correction and Deformable Registration in Dynamic Perfusion Imaging

Dynamic myocardial CTP acquisitions are vulnerable to motion artifacts arising from cardiac contraction, respiration, and involuntary patient movement [19,20]. Deformable registration is the standard motion-correction strategy in many CT and MR applications, as rigid and affine transforms cannot capture complex, nonlinear tissue motion. Broadly, deformable methods can be divided into optimization-based (classical) algorithms [22,60] and learning-based approaches [24,61].

Optimization-based techniques, including optical flow, B-splines, and diffeomorphic registration, minimize an energy functional combining image similarity and deformation regularity. In lung CT, publicly available datasets such as DIR-Lab have enabled significant advances and benchmarking [62,63]. In cardiac MR perfusion, Benovoy et al. [23] proposed a nonrigid registration framework that exploits proton density cues and temporal optical flow, aligning frames using Histogram of Oriented Gradients (HoG) features between consecutive images. Their approach improved correlation metrics in cropped cardiac regions, though quantitative TRE evaluation was not possible due to missing landmarks. Huang et al. [64] introduced a reference-free, learning-based similarity metric for motion compensation in cone-beam CT, eliminating the need for ground-truth motion fields by learning image similarity directly from data. While these works highlight robust strategies for motion compensation, they are not specifically tailored to dynamic myocardial CTP and often operate on static or quasi-static datasets.

Deep learning–based registration methods have emerged as a complementary paradigm, learning parameterized deformation fields directly from image pairs or sequences. VoxelMorph-type architectures [24,65], multi-scale unsupervised frameworks [62,66], and modality-specific networks [25,26,67] have demonstrated competitive accuracy and efficiency in lung, abdominal, and neurosurgical imaging. Fechter and Baltas [68] developed a one-shot multi-scale CNN with parallel U-Nets capturing small, medium, and large deformations, evaluated primarily on lung datasets. Lara-Hernandez et al. [69] proposed a deep learning registration framework for dynamic myocardial perfusion CT imaging, and other modality-focused works have investigated registration for PET and SPECT myocardial perfusion imaging [27,28,70]. Nonetheless, the lack of annotated myocardial CTP datasets with ground-truth deformation fields, combined with rapid contrast changes that violate brightness constancy, complicates both training and validation [24,61,71].

Clinical studies underscore the importance of motion correction but often rely on conventional tools or manual alignment, rather than introducing new algorithms. Mooiweer et al. [20] implemented a fast navigator strategy for prospective respiratory motion correction in MR perfusion, and Otaki et al. [19] combined motion correction with residual activity correction in PET myocardial perfusion imaging. Hubbard et al. [21] proposed a motion-immune (MI) strategy for dynamic myocardial CTP that avoids registration entirely; in a swine model, they reported high MBF correlation and improved detection of perfusion deficits compared to motion-susceptible methods. Despite these advances, there remains a paucity of motion-correction frameworks explicitly designed, trained, and quantitatively validated for dynamic myocardial CT perfusion, particularly in low-dose regimes and 4D settings.

Motion degradation in low-dose myocardial CTP was previously addressed using a modified one-shot CNN–based registration model [72]. In that work, the one-shot architecture was extended to a spatiotemporal (4D) formulation by combining slice-wise deformation estimation with B-spline interpolation across time, and was further guided by a TEC-consistency loss term, ${\mathcal{L}}_{\mathrm{TEC}}$, to preserve temporal enhancement behaviour better. The reported results demonstrated effective motion correction in low-dose myocardial CTP, motivating the use of this motion-correction model as one of the core components within the present unified motion and noise correction framework.

2.3. Other Approaches

While most prior work treats noise suppression and motion correction as separate preprocessing steps, their interaction is nontrivial: aggressive denoising can remove subtle structural cues needed for accurate registration, whereas motion correction can distort intensity patterns critical for noise modelling and perfusion quantification. Accordingly, there is growing interest in approaches that embed denoising or motion correction within a perfusion-specific objective.

Some methods operate at the level of perfusion maps rather than individual time frames, effectively coupling denoising with perfusion quantification. Kadimesetty et al. [57] used CNNs to denoise CT perfusion maps directly, sidestepping explicit preprocessing but sacrificing frame-level control. Ramon et al. [58] trained a 3D convolutional autoencoder to predict high-dose SPECT myocardial perfusion images from low-dose acquisitions, thereby integrating noise suppression, partial volume recovery, and perfusion enhancement in a single network. Some imaging studies have combined motion correction with integrated pipelines for SPECT and PET [27,28,70], though these are typically modality-specific and do not address the mixed Poisson–Gaussian noise and rapid contrast dynamics of myocardial CTP. One specific work, conducted by Lukas et al. [73], observed denoising as a by-product while emphasizing motion correction.

In the CT domain, recent trends include hybrid loss functions that integrate pixel-wise, structural, and perceptual terms; multi-scale architectures that capture both global and local structure; and transformer-based or diffusion-based generative models that can be dose-aware or zero-shot adaptable [74,75,76,77,78]. However, most of these frameworks target static or quasi-static CT, and their application to fully dynamic myocardial CTP with explicit motion modelling remains limited. Motion-focused works have begun to incorporate perfusion-specific cues, such as time-enhancement curve consistency or region-of-interest constraints [26,69], but few explicitly unify noise modelling, deformable motion correction, and perfusion-aware optimization within a single architecture tailored to low-dose myocardial CTP.

While joint frameworks that simultaneously address motion and image quality have been explored in other dynamic imaging contexts [79], such approaches remain relatively limited for myocardial CT perfusion (CTP), particularly for image-domain post-processing of reconstructed sequences. Existing CTP studies predominantly focus on either spatio-temporal restoration or denoising without explicit motion correction [80,81]. In contrast, unified motion-denoising methods are typically developed for different modalities, such as PET or MRI [79,82], or operate at the reconstruction level rather than on reconstructed images. We would also like to highlight the ASTRA4D algorithm, which has been clinically evaluated for four-dimensional noise reduction and motion elimination in low-dose dynamic myocardial CT perfusion, demonstrating improved image quality and diagnostic sensitivity compared with temporal PCA-based approaches [73]. While this represents a joint noise reduction and motion compensation method for myocardial CTP, it differs from our approach in that it does not employ learned models or a unified deep image-domain pipeline, which is the focus of our proposed framework. These differences in modality, processing domain, and formulation make direct comparison challenging. In this work, we therefore propose a unified image-domain framework tailored for low-dose myocardial CTP, integrating denoising and motion correction into a single workflow and evaluating its impact under simulated and real low-dose conditions.

In summary, existing literature offers: (i) powerful LDCT denoising methods, mainly developed and validated on non-cardiac datasets; (ii) a rich ecosystem of deformable registration techniques, with limited myocardial CTP–specific validation; and (iii) early attempts at integrating denoising/motion correction and perfusion analysis, primarily in SPECT/PET or at the perfusion-map level. Except Lukas et al. [73], to the best of our knowledge, no prior work has introduced a unified framework that simultaneously models mixed Poisson–Gaussian noise and dynamic cardiac motion in low-dose myocardial CTP, while being evaluated with both image quality and motion-specific metrics such as TEC consistency and TRE. This gap motivates the combined noise–motion framework proposed in this study.

3. Problem Statement

Despite substantial progress in both noise suppression and motion correction, a clear gap persists in low-dose dynamic myocardial CTP imaging. Most denoising studies focus on general low-dose CT or non-cardiac regions [48], with limited attention to myocardial perfusion imaging, where rapid contrast dynamics and cardiac motion introduce unique challenges [21,83]. Likewise, many motion correction techniques have been developed for static CT or MR perfusion [84], often relying on synthetic datasets or simplified motion models that do not fully capture the interplay of cardiac contraction, respiration, and contrast flow. Few studies explicitly consider how noise influences motion estimation and how motion, in turn, affects denoising performance, leaving an unsolved problem of jointly addressing both degradations in low-dose myocardial CTP [27].

The dual degradations in low-dose myocardial CTP, noise and motion artifacts are tightly coupled. Motion artifacts reduce the temporal coherence of voxel-wise intensity curves, while noise obscures anatomical boundaries that registration algorithms rely upon. Conventional sequential pipelines, which apply denoising followed by registration (or vice versa), implicitly assume that the two problems can be treated independently. In practice, denoising prior to registration risks oversmoothing or distorting structural gradients, which are essential for accurate motion estimation [85]. In contrast, motion correction prior to denoising may misinterpret noisy fluctuations as anatomical features, leading to erroneous deformation fields [65,66]. The central research challenge is therefore to design a framework that robustly corrects both noise and motion jointly, preserving fine myocardial structures, ensuring realistic and consistent motion fields, and maintaining robustness under low-dose CTP conditions where both degradations are pronounced [77,78].

3.1. Informal Problem Description

Given a low-dose dynamic CTP series, the goal of this work is to reconstruct a CT perfusion image sequence that is (i) motion-consistent across time and (ii) noise-suppressed without blurring diagnostically relevant myocardial structures. Achieving both objectives is essential for downstream perfusion parameters (e.g., myocardial blood flow) to be estimated with accuracy comparable to that obtained from high-dose, motion-free acquisitions.

Traditional sequential pipelines treat noise and motion as separate problems, performing denoising followed by registration, or vice versa. However, such approaches fail to account for the strong coupling between the two degradations: excessive noise makes motion fields unreliable, while misregistration can distort anatomical detail and complicate subsequent denoising. This motivates a new strategy in which motion correction and noise suppression are addressed in a coordinated manner rather than as isolated, serial operations.

The proposed solution is to unify two deep learning models, each specialized for a different type of degradation, and then combine their complementary strengths via an ensemble. A modified one-shot CNN is trained to capture motion patterns across dynamic frames, yielding a motion-consistent candidate sequence. In parallel, a DnCNN-inspired, modified FFDNet architecture is trained to learn the low-dose noise distribution, generating a denoised candidate sequence. The final output is obtained by an ensemble fusion step, where the two candidates are blended through weighted averaging.

By unifying these tasks in a parallel ensemble framework, the system aims to ensure that neither correction compromises the other. Motion correction preserves temporal alignment, which is essential for perfusion analysis, while denoising enhances the visibility of fine myocardial features. Their combination produces high-quality, temporally coherent CT perfusion sequences that preserve image characteristics relevant to quantitative evaluation in clinical practice.

3.2. Formal Problem Definition

A low-dose dynamic myocardial CTP series can be represented as:

$$\mathcal{X}={\left\{x_{z,t}\right\}}_{z=1,\dots ,Z;t=1,\dots ,T}$$

which forms a 4D tensor $\mathcal{X}\in {\mathbb{R}}^{H\times W\times Z\times T}$, where $H\times W$ are the in-plane dimensions, Z is the number of slices, and T is the number of time frames. The target is to reconstruct the motion-consistent, noise-suppressed series of the same dimensions.

$$\mathcal{Y}={\left\{y_{z,t}\right\}}_{z=1,\dots ,Z;t=1,\dots ,T}$$

The observed slices are noisy, deformed versions of the underlying clean 3D volume at each time point. Let $\varphi$ denote a 4D deformation field defined over the full volume, capturing nonrigid motion across space and time, and the corresponding warp operator be ${\mathcal{W}}_{\varphi }$. The inverse problem of recovering motion-consistent and noise-suppressed 4DCTP data can be conceptualized as

$$\mathcal{Y}={\left[\mathcal{N}\circ {\mathcal{W}}_{\varphi }\left(\mathcal{X}\right)\right]}_{z,t}$$

(1)

where ${\mathcal{W}}_{\varphi }$ applies the 4D motion warp, $\mathcal{N}(·)$ models low-dose noise corruption, and ${[·]}_{z,t}$ extracts slice z at time t. Here $\mathcal{N}\circ {\mathcal{W}}_{\varphi }$ is interpreted as a compound degradation operator in which both motion and noise contribute simultaneously to the formation of the measured data.

3.3. Learning Objective

Assuming training data ${\left\{({\mathcal{X}}^{\left(i\right)},{\mathcal{Y}}^{★\left(i\right)})\right\}}_{i=1}^N$ (paired or synthetically paired, where $\mathcal{X}$ denotes degraded low-dose CTP and ${\mathcal{Y}}^{★}$ denotes corresponding high-dose CTP), the ensemble parameters $\Theta =(\theta ,\psi ,w)$ are optimized to minimize a composite risk:

$$\underset{\theta ,\psi ,w}{min}\mathbb{E}\left[{\mathcal{L}}_{\mathrm{img}}(\mathcal{Y},{\mathcal{Y}}^{★})+{\lambda }_{\mathrm{mc}}{\mathcal{L}}_{\mathrm{motion}}\left({\mathcal{Y}}^{\mathrm{motion}}\right)+{\lambda }_{\mathrm{ns}}{\mathcal{L}}_{\mathrm{noise}}\left({\mathcal{Y}}^{\mathrm{noise}}\right)\right].$$

(2)

Here:

• $\theta$ are the learnable parameters of the motion correction network ${\mathcal{M}}_{\theta }$,
• $\psi$ are the learnable parameters of the noise correction network ${\mathcal{D}}_{\psi }$,
• w are the ensemble fusion weights that combine the outputs of the two models,
• ${\lambda }_{\mathrm{mc}},{\lambda }_{\mathrm{ns}}>0$ are hyperparameters that control the relative influence of motion-regularization and structural-preservation penalties during training.

The three terms capture complementary requirements:

• ${\mathcal{L}}_{\mathrm{img}}$ promotes fidelity between the final output $\mathcal{Y}$ and the ideal target ${\mathcal{Y}}^{★}$.
• ${\mathcal{L}}_{\mathrm{motion}}$ encourages temporal coherence and smooth, physically plausible deformations in the motion-corrected candidate ${\mathcal{Y}}^{\mathrm{motion}}$.
• ${\mathcal{L}}_{\mathrm{noise}}$ penalizes the loss of fine anatomical details, ensuring that diagnostically relevant edges and myocardial features are retained in ${\mathcal{Y}}^{\mathrm{noise}}$.

It is important to note that these terms are not added qualities in a physical sense. Instead, their weighted sum serves as a training objective: the $\lambda$ values determine how strongly the optimization prioritizes motion consistency, structural fidelity, or image similarity. In contrast, the ensemble weight w governs how the outputs of the motion and noise models are combined at inference time to form the final prediction.

3.4. Scope and Assumptions

This work targets image-domain motion and noise correction for contrast-enhanced dynamic myocardial CTP. We assume that:

• Low-dose noise after standard CT reconstruction can be approximated by a mixed Poisson–Gaussian model, which is handled via variance-stabilizing transformations and learned denoising, as demonstrated in [59].
• Deformations between cardiac phases are smooth and invertible within the myocardium, allowing representation by a 4D deformation field with physiologically plausible regularity [72].
• Scanner-specific reconstruction algorithms are treated as fixed; variations in dose, heart rate, and acquisition protocol are addressed through data augmentation and appropriately designed loss terms.

The proposed framework is agnostic to the specific perfusion modelling technique used downstream. It can be integrated with conventional voxel-wise time–attenuation curve analysis or other quantitative perfusion estimation methods. Its primary goal is to provide motion-consistent, noise-suppressed 4D CTP images that form a robust foundation for accurate myocardial perfusion quantification under low-dose imaging conditions.

4. Proposed Method

Deep learning models offer a promising direction for addressing the combined noise–motion artifact challenge. Unlike traditional filters or iterative registration pipelines, neural networks can learn complex nonlinear mappings directly from data, thereby capturing the inherent characteristics of noise and motion [74]. To overcome the research gaps and limitations identified in myocardial CTP, we propose a unified deep learning framework that ensembles the outputs of two parallel models that target Poisson–Gaussian noise and temporal motion. This approach improves low-dose noise suppression while preserving diagnostically relevant perfusion features and temporal dynamics.

4.1. Noise Correction Model (Prior Work)

Noise degradation in low-dose myocardial CTP has been previously addressed using an extended FFDNet-based denoising framework [59]. FFDNet builds upon the DnCNN architecture by incorporating a noise-level map as an additional input, allowing adaptive handling of spatially varying noise [14,50]. This architecture was selected for its ability to incorporate explicit noise-level information and maintain stable performance across varying dose conditions. While the original FFDNet formulation assumes additive Gaussian noise, this assumption does not adequately capture the mixed Poisson–Gaussian noise statistics inherent to CT imaging [54]. To address this mismatch, the prior work introduced a Generalized Anscombe Transform (GAT) to stabilize signal-dependent Poisson noise into an approximately signal-independent Gaussian distribution [59,86], enabling effective learning under FFDNet’s Gaussian-noise assumption [85].

The impact of incorporating GAT was systematically evaluated in an ablation study reported in [59]. The results showed that DnCNN/FFDNet without GAT exhibited limited denoising performance due to a mismatch between the underlying Poisson–Gaussian noise characteristics of CT data and the network’s Gaussian noise assumption. In contrast, applying GAT prior to network inference consistently resulted in substantially lower MSE, higher PSNR, and improved SSIM across all evaluated test cases. These findings indicate that variance stabilization plays a critical role in enabling effective denoising of low-dose myocardial CT perfusion images when using Gaussian-noise–oriented deep learning architectures.

For completeness and reproducibility, we briefly summarize the denoising network’s training configuration. The model follows a DnCNN-inspired FFDNet architecture consisting of sequential $3\times 3$ convolutional layers with batch normalization and ReLU activation, augmented with dilated convolutions for multi-scale feature extraction and residual learning to predict the noise component in the variance-stabilized domain. The network input comprises the noisy image (after GAT) and an associated noise-level map. The loss function is mean squared error (MSE) computed between the predicted and high-dose reference images in the transformed domain. Training was performed using stochastic gradient descent (SGD) with an initial learning rate of ${10}^{-3}$ and momentum of 0.8, a batch size of 16, and early stopping with a tolerance of 8 epochs (convergence occurred after 12 epochs, with a maximum of 50 epochs allowed). The dataset included 8960 image pairs derived from four CTP studies, with 7960 used for training and 1000 for validation. Default FFDNet downsampling was employed. The average inference time was approximately 1.2 s per image on an Intel Core i5-4570 CPU (3.20 GHz), with GPU-based training on NVIDIA hardware.

This denoising model serves as the noise-correction backbone in the proposed unified framework.

4.2. Motion Correction Model (Prior Work)

The motion correction component employed in the present framework has been previously developed and validated in [72]. That work builds upon a one-shot CNN registration paradigm [68], which predicts deformation fields in a single forward pass, offering substantial computational advantages over iterative deformable registration approaches [87]. To adapt the one-shot framework to dynamic myocardial CT perfusion, the model was extended to a spatiotemporal (4D) formulation by combining slice-wise deformation estimation with B-spline interpolation across time, resulting in a single continuous 4D deformation vector field. B-spline refinement was used to enforce spatial smoothness and anatomical plausibility of the estimated deformation fields. In addition, a time–enhancement-curve (TEC)-guided consistency term was incorporated into the training objective to preserve temporal enhancement behaviour across frames explicitly. The resulting model demonstrated effective correction of cardiac and respiratory motion in low-dose myocardial CTP.

For reproducibility, we briefly summarize the training configuration of the motion-correction model. Physiologically plausible motion was simulated by applying random nonrigid deformations with a maximum displacement of 3 pixels (approximately 2 mm, consistent with the dataset’s voxel resolution) to dynamic CTP volumes. Each original–deformed pair served as fixed and moving inputs to the one-shot CNN, which was optimized directly on this pair without batch-based dataset training. The network follows a multi-scale encoder–decoder architecture, progressively downsampling inputs to capture motion at multiple spatial resolutions and predicting a deformation field in a single forward pass. The loss function consists of three components: (i) anatomical similarity between fixed and warped images, (ii) temporal consistency enforced via the TEC-guided term ${\mathcal{L}}_{\mathrm{TEC}}$, and (iii) deformation-field smoothness regularization. Slice-wise deformation fields were subsequently interpolated using cubic B-spline interpolation to generate a unified 4D spatiotemporal deformation field across all frames. This training strategy enables motion estimation that respects both anatomical alignment and perfusion dynamics while avoiding iterative optimization.

In the present work, this published motion-correction model is treated as a fixed, validated module and integrated with a dedicated denoising component within a unified ensemble framework. Detailed architectural design, training strategy, and quantitative evaluation are provided in [72].

4.3. Combined Artifact Correction

Finally, we propose an ensemble strategy that integrates the outcomes of the noise and motion correction models. Instead of applying denoising and registration sequentially, which risks degrading features required by the other, our approach leverages both corrections through a unified inference step. The denoised image from the modified FFDNet and the motion field from the improved one-shot registration network are fused via a consensus mechanism. The fusion strategy employs a weighted averaging scheme, chosen for its stability, interpretability, and preservation of temporal consistency without additional trainable parameters. An iterative coupling strategy and more complex fusion mechanisms were not explored in order to avoid increased model complexity and potential error accumulation. This ensemble approach enables the models to compensate for each other’s limitations, performing a unified inference to yield motion-corrected, denoised myocardial CTP images that preserve anatomical fidelity and enhance the reliability of perfusion quantification. Figure 1 illustrates the proposed architecture.

4.4. Training and Testing

Experiments were conducted using anonymized porcine cardiac CT perfusion (CTP) datasets acquired at three tube current settings: 80 mA (high dose), 40 mA (mid dose), and 20 mA (low dose). The dataset comprised 16 studies in total, including data from prior model development and independent studies reserved for evaluating the unified framework across different dose levels. As detailed in [59,72], realistic low-dose noise characteristics and physiologically plausible cardiac and respiratory motion were modelled during the development of the individual denoising and motion-correction components, using noise-stabilization strategies and perfusion-aware objectives to preserve both anatomical structure and temporal enhancement behaviour. In the present study, these previously trained components are evaluated and integrated within a unified framework without further modification.

The denoising component employed in the present framework was previously developed and validated in [59]. In that work, mixed Poisson–Gaussian noise was simulated and stabilized using a Generalized Anscombe Transform, with noise-level maps provided as auxiliary inputs to an extended FFDNet architecture. The model was trained on low-dose myocardial CTP data using a controlled training and validation protocol, and details regarding data preparation, noise simulation, network architecture, and training strategy are reported therein. The reader is referred to [59] for a complete methodological description.

The motion correction component employed in the present framework was previously developed and validated in [72]. In that work, motion was modelled using a modified one-shot CNN registration paradigm, in which deformation fields were estimated from fixed–moving volume pairs without reliance on batch-based training. Slice-wise deformation estimates were subsequently integrated into a continuous spatiotemporal deformation field using B-spline interpolation to represent motion across the full dynamic sequence. Full details regarding motion simulation, training strategy, deformation field construction, and optimization are reported in [72], and are not repeated here.

The unified framework was evaluated on a total of 16 myocardial CT perfusion datasets, comprising 8 high-dose (80 mA) and 8 low-dose (20 mA or lower) acquisitions. For the high-dose datasets, corresponding simulated low-dose images were generated, enabling reference-based image quality assessment by direct comparison with high-dose references after motion and noise correction. For datasets acquired directly under low-dose protocols, where no high-dose reference was available, performance was assessed using no-reference image quality metrics. In addition to quantitative evaluation, expert reader assessment was conducted using a predefined scoring matrix that evaluated multiple aspects of perceived image quality on a five-point scale (1–5).

4.5. Implementation and Integration Details

In the unified framework, both the denoising and motion-correction models were used with their pretrained weights fixed, and no additional joint fine-tuning or end-to-end backpropagation was performed. Weighted averaging was selected for ensemble operation due to its stability and interpretability, and because preliminary experiments indicated that more complex fusion strategies, such as iterative coupling or learned fusion layers, did not provide consistent improvements while significantly increasing model complexity. The ensemble fusion weight w was determined empirically by evaluating performance across multiple candidate settings and selecting the configuration that yielded the most stable quantitative and qualitative results. To assess robustness, preliminary sensitivity experiments were conducted for $w\in [0.4,0.7]$. Within this interval, the unified framework exhibited consistent performance trends without qualitative degradation or substantial quantitative variation, indicating that the integration is not critically dependent on a single finely tuned weight value. Accordingly, the reported configuration was adopted for consistency across experiments. The integration strategy, therefore, relies on validated component-level training followed by empirical fusion rather than combined optimization across both branches. Because no additional trainable parameters or iterative coupling were introduced, the inference time remains effectively comparable to that of the denoising module alone. Furthermore, since no joint training was conducted at the ensemble level, no additional data splitting or cross-validation procedure was required beyond those already implemented in the original component-specific studies.

5. Experimental Setup and Results

This section summarizes the datasets, experimental protocol, and quantitative and qualitative evaluation of the proposed unified noise–motion correction framework for low-dose dynamic myocardial CTP. Results are reported for both simulated and real low-dose acquisitions, and for three task settings: denoising only, motion correction only, and combined degradation.

5.1. Datasets

Experiments were conducted on 16 porcine myocardial CTP studies from 6 animals, acquired at multiple tube currents ranging from $80\mathrm{mA}$ (reference) down to $10\mathrm{mA}$ (ultra-low dose).

Table 2. Summary of Test Cases: Study Identifiers, Radiation Dose Levels, and Number of Images.

Study Identifier	Dose	Number of Images
STC1	80 mA	2376
STC2	80 mA	304
STC3	80 mA	392
STC4	80 mA	176
STC5	80 mA	1120
STC6	40 mA	1120
STC7	80 mA	1120
STC8	40 mA	1120
RTC1	20 mA	2376
RTC2	20 mA	368
RTC3	20 mA	400
RTC4	20 mA	176
RTC5	20 mA	1120
RTC6	20 mA	1120
RTC7	15 mA	1120
RTC8	10 mA	1120

summarizes the datasets.

High-dose scans (80 mA) were used to construct Simulated Test Cases (STC), where realistic noise and motion were synthetically introduced under controlled conditions. These enable reference-based evaluation against the original high-dose images. Low-dose scans acquired in practice constitute Real Test Cases (RTC) and are used for no-reference and expert-based evaluation. All studies, both low-dose and high-dose, used in this work are from real animal experiments (no phantoms).

5.2. Image Quality Assessment (IQA) Metrics

To evaluate the performance of the proposed artifact-correction framework, we use a combination of reference-based and no-reference IQA metrics. Reference-based metrics compare the processed image with a high-dose ground truth, enabling direct quantification of noise suppression and structural fidelity. However, in myocardial CTP imaging, high-dose reference data are often unavailable or impractical to acquire due to radiation safety concerns. In such cases, no-reference metrics are essential, as they assess perceptual image quality without relying on ground-truth images. This allows evaluation of denoising and motion correction performance in clinically relevant scenarios where only low-dose images are available. The benefit of using no-reference IQA lies in its ability to generalize across real-world conditions, providing an unbiased measure of diagnostic quality in the absence of a standard reference.

By combining reference-based and no-reference metrics, we can rigorously assess denoising and motion correction performance. Reference-based measures (MSE, PSNR, SSIM, TRE) provide quantitative validation against simulated high-dose images, while no-reference metrics (FID, KID, NIQE, Noise Variance, Correlation Coefficient) ensure evaluation under practical low-dose conditions where ground truth is not accessible. In addition to reference-based and reference-less IQAs, we evaluated our performance through visual inspection, comprehensive expert opinion, and quantitative assessments, including the Area Under the Curve (AUC). This multidimensional evaluation strategy offers a thorough assessment of image quality improvements, both objectively and in clinically relevant contexts.

The evaluation protocol is organized along two axes: (i) simulated versus real low-dose data and (ii) task type: denoising, motion correction, and combined noise and motion correction (

Table 3. Summary of Testing Strategy vs. Task Type.

Testing Type	Task	Evaluation Process & Metrics
Simulated Testing	Denoising	Add known noise; measure MSE, PSNR and SSIM after denoising.
	Motion Correction	Add known deformation; measure TRE and Correlation Coefficient.
	Combined	Add both noise and deformation; Calculate MSE, PSNR, SSIM, FID, KID, and NIQE.
Real Testing	Combined	Calculate FID, KID, Noise Variance and NIQE; Analyze AUC and Correlation Coefficients; Obtain radiologists’ scoring.

).

5.2.1. Simulated Studies (Reference-Based)

For simulated experiments, the 80 mA images serve as reference, and the following tasks are defined:

• Denoising:Poisson–Gaussian noise is applied to 80 mA images to emulate 20 mA acquisitions. The proposed denoising network reconstructs high-quality images, and evaluated by MSE, PSNR, and SSIM scores.
• Motion Correction: Known non-rigid deformations (up to 3 pixels) are applied to simulate cardiac/respiratory motion. The motion network registers deformed images back to the reference, primarily evaluated by Target Registration Error (TRE) and the correlation of Time Enhancement Curves (TECs).
• Combined Degradation: Both noise and motion are applied sequentially to the reference images. The complete unified model (motion and noise branches with ensemble fusion) is evaluated using MSE, PSNR, SSIM, TRE and no-reference metrics (FID, KID, NIQE).

5.2.2. Real Low-Dose Studies (No-Reference)

For RTC1–RTC8, only noisy low-dose reconstructions are available. The unified model is applied directly to these series and evaluated using:

• No-reference metrics (FID, KID, NIQE and noise variance),
• Quantitative functional consistency (AUC and Pearson correlation of TECs),
• Comprehensive subjective ratings by experts.

Where available, high-dose reference scans from matching studies are used to correlate AUC and TEC in selected regions (e.g., the left superior pulmonary vein and the descending aorta).

5.3. Objective Image Quality Assessment

For the simulated test cases (STC1–STC8), we added synthetic noise and motion to evaluate the noise and motion correction models, respectively, using reference-based image quality assessments. After applying simulated noise, we calculated their FID and KID scores to confirm that the synthetic low-dose images closely resemble real low-dose scans, as shown in

Table 4. Comparison of FID and KID values between actual low-dose scan and simulated noisy images. The closeness in the pair indicates the simulation resembles the low-dose scan.

Test Case	Scan Dose/ Sim. Dose	Low Dose Scan			Low Dose Simulation
			KID			KID
		FID	Mean	Std. Dev.	FID	Mean	Std. Dev.
RTC1/STC1	20/20 mA	1.0670	1.0050	0.9504	1.4841	1.1614	0.4205
RTC2/STC2	20/20 mA	1.1794	1.0852	0.9477	1.5997	1.3145	0.6425
RTC3/STC3	20/20 mA	1.1763	1.1542	1.1247	1.5973	1.2368	0.5226
RTC4/STC4	20/20 mA	1.1545	1.1293	1.0323	1.4461	1.2320	0.5512
RTC5/STC5	20/20 mA	2.6024	4.6036	2.6247	1.7506	1.5137	0.5176
RTC6/STC6	20/20 mA	3.1834	5.9372	3.5163	2.2662	2.4250	0.7242
RTC7/STC7	15/20 mA	5.3388	10.1893	5.4318	2.2332	2.1443	0.7737
RTC8/STC8	10/20 mA	10.9367	21.8039	8.9176	3.0284	3.8554	1.5322

. This validates that our simulation resembles the actual low-dose scans.

5.3.1. Evaluating Noise Correction Model (Prior Work)

The noise correction component used in the present unified framework was previously developed and extensively evaluated in [59]. That study quantitatively assessed denoising performance on simulated low-dose myocardial CTP data using reference-based image quality metrics, including MSE, PSNR, and SSIM, demonstrating substantial noise reduction and structural preservation relative to high-dose references. In addition, the impact of variance stabilization via the Generalized Anscombe Transform was systematically analyzed, showing that stabilizing mixed Poisson–Gaussian noise into an approximately Gaussian distribution is critical for effective learning with Gaussian-noise–oriented architectures such as DnCNN and FFDNet. Comparative evaluations against recent CNN, GAN, and transformer-based LDCT denoising methods further confirmed the competitiveness of the proposed approach under realistic perfusion-specific noise conditions. For completeness, the reader is referred to [59] for detailed quantitative results, ablation studies, and comparative analyses. In the present work, this denoising model is treated as a fixed, validated module and integrated into a unified framework with motion correction.

5.3.2. Evaluating Motion Correction Model (Prior Work)

The motion correction component integrated into the present unified framework was previously developed and extensively evaluated in [72]. That study assessed registration performance on dynamic myocardial CT perfusion data using anatomically meaningful landmarks and reference-based metrics, including target registration error (TRE), demonstrating substantial reduction of nonrigid cardiac and respiratory motion. In addition to spatial alignment accuracy, temporal consistency was evaluated through correlation analysis of time–enhancement curves (TECs), confirming effective preservation of perfusion dynamics following motion correction. Comparative analyses against established deformable registration methods reported in the literature further showed that the proposed approach achieves competitive TRE values despite the added challenges of dynamic contrast enhancement and simultaneous cardiac and respiratory motion in myocardial CTP. For completeness, the reader is referred to [72] for detailed quantitative results, landmark definitions, visual assessments, and comparative evaluations. In the present work, this motion-correction model is treated as a fixed, validated module and integrated with a denoising component within a unified ensemble framework.

5.3.3. Compound Noise and Motion Correction

When both degradations are present, the unified model still achieves substantial improvements. For compound simulations (STC1–STC8), MSE drops from approximately 33–44 to 5–7 (

Table 5. Mean Squared Errors (MSE) calculated from Compound Simulated Reconstructed Test Cases with 95% Confidence Intervals.

Test Case	MSE Noisy				MSE Recon. (80 mA)
	Mean	Min	Max	Std. Dev.	Mean	Min	Max	Std. Dev.	95% CI (±)
STC1	44.13	35.76	584.13	12.97	6.79	2.31	23.93	3.35	0.14
STC2	43.36	35.49	544.40	31.03	5.62	2.16	22.18	3.25	0.37
STC3	44.24	37.13	582.06	29.25	6.78	2.51	32.60	4.16	0.41
STC4	41.97	33.85	532.91	39.74	6.41	2.49	16.85	3.28	0.49
STC5	36.31	29.91	519.41	15.83	5.31	3.20	19.93	2.05	0.12
STC6	35.52	30.43	521.11	15.94	6.78	4.07	27.24	2.66	0.16
STC7	33.61	27.30	497.45	15.40	5.39	3.26	22.16	2.36	0.14
STC8	35.31	27.36	483.94	15.33	7.07	4.31	30.90	2.90	0.17

), and PSNR increases from ∼31.8–32.8 dB to ∼39.9–41.2 dB (

Table 6. Peak Signal to Noise Ratio (PSNR) calculated from Compound Simulated Reconstructed Test Cases with 95% Confidence Intervals.

Test Case	PSNR Noisy				PSNR Recon. (80 mA)
	Mean	Min	Max	Std. Dev.	Mean	Min	Max	Std. Dev.	95% CI (±)
STC1	31.74	20.47	32.60	0.58	40.30	34.34	44.50	2.05	0.08
STC2	31.96	20.77	32.63	0.90	41.22	34.67	44.78	2.19	0.25
STC3	31.83	20.48	32.43	0.81	40.43	33.00	44.14	2.23	0.22
STC4	32.23	20.86	32.83	1.12	40.58	35.86	44.16	2.11	0.31
STC5	32.53	20.98	33.37	0.58	41.13	35.13	43.08	1.41	0.08
STC6	32.62	20.96	33.30	0.59	40.07	33.78	42.03	1.40	0.08
STC7	32.96	21.16	33.77	0.67	41.11	34.67	43.00	1.51	0.09
STC8	32.75	21.28	33.76	0.74	39.90	33.23	41.79	1.42	0.08

). SSIM improves from ∼0.54–0.59 (heavily degraded inputs) to ∼0.89–0.91 (

Table 7. Structural Similarity Index (SSIM) calculated from Compound Simulated Reconstructed Test Cases with 95% Confidence Intervals.

Test Case	SSIM Noisy				SSIM Recon. (80 mA)
	Mean	Min	Max	Std. Dev.	Mean	Min	Max	Std. Dev.	95% CI (±)
STC1	0.54	0.29	0.58	0.02	0.91	0.91	0.91	0.00	0.00008
STC2	0.55	0.31	0.57	0.02	0.91	0.91	0.92	0.00	0.00023
STC3	0.54	0.27	0.56	0.02	0.91	0.91	0.91	0.00	0.00020
STC4	0.57	0.31	0.59	0.03	0.91	0.91	0.91	0.00	0.00030
STC5	0.58	0.26	0.60	0.02	0.90	0.90	0.90	0.00	0.00012
STC6	0.57	0.26	0.59	0.02	0.89	0.89	0.90	0.00	0.00012
STC7	0.59	0.28	0.62	0.02	0.90	0.89	0.91	0.00	0.00018
STC8	0.58	0.28	0.61	0.02	0.89	0.88	0.89	0.00	0.00018

). Across all STC cases, statistical comparison between noisy and reconstructed images yielded two-sided p-values below the reporting threshold ($p<0.0001$) for MSE, PSNR, and SSIM. This consistency reflects the combination of large sample sizes (n = 176–2376 per case), which reduces the standard error, and substantial effect magnitudes observed across all metrics. These factors yield high-magnitude test statistics in each case, leading to uniformly small p-values. Accordingly, p-values are reported as $p<0.0001$, as more precise numerical values would not provide additional interpretive insight. The statistical significance is further supported by narrow 95% confidence intervals and clinically meaningful effect sizes.

TRE values in the compound setting remain low, with mean errors ranging from 1.89 to 3.13 pixels (

Table 8. Target Registration Error (TRE) calculated from Compound Simulated Reconstructed Test Cases.

Test Case	Target Reg. Error
	Mean	Min	Max	Std. Dev.
STC1	3.13	0.00	8.06	3.35
STC2	2.17	1.00	4.12	1.23
STC3	1.89	1.00	4.24	1.29
STC4	2.80	0.00	8.06	3.00
STC5	2.29	1.00	3.61	1.18
STC6	2.80	1.41	4.24	1.02
STC7	2.24	1.00	6.40	1.90
STC8	2.06	0.00	5.10	1.55

), consistent with the deformation range used during training and testing. This confirms that the unified model can jointly correct noise and motion while preserving spatial accuracy.

5.3.4. No Reference Image Quality Assessments

To complement reference-based metrics, we employed no-reference image quality assessment (NR-IQA) measures, including Fréchet Inception Distance (FID), Kernel Inception Distance (KID), and NIQE. These metrics were used to evaluate perceptual and structural consistency in settings where a paired high-dose reference may not be available, such as real low-dose clinical acquisitions [59].

We acknowledge that standard implementations of FID and KID rely on feature representations derived from ImageNet-trained networks, which were originally developed for natural scene images. To mitigate potential domain mismatch, we computed feature-space statistics using high-dose CT images from our dataset rather than relying on generic ImageNet population statistics. In addition, NIQE parameters were re-estimated using the same high-dose CT dataset to construct a domain-adapted natural scene statistics model tailored to myocardial CT perfusion imaging. This approach allows the no-reference metrics to operate within a CT-specific statistical feature space rather than a natural-image prior.

Importantly, FID, KID, and NIQE are not interpreted as stand-alone clinical validation tools but as complementary statistical descriptors of feature-space similarity, smoothness, and structural consistency. Their inclusion is intended to provide additional evidence of distributional alignment between reconstructed and high-dose images, alongside established reference-based metrics (MSE, PSNR, SSIM) and temporal consistency measures. While the use of perceptual feature-space metrics in CT perfusion imaging remains an area of ongoing investigation, our results demonstrate consistent trends across both reference-based and no-reference evaluations along with experts’ subjective scoring, supporting the stability of the proposed framework. A comprehensive validation of these perceptual metrics with respect to downstream diagnostic tasks is beyond the scope of this study and is identified as future work.

We observed that after reconstruction, FID consistently decreases for real RTC tests (

Table 9. FID scores for noisy and reconstructed CT images calculated across different test cases.

Test Case	Dose/ Tube Current	FID
		Noisy	Recon. (80 mA)
RTC1	20 mA	1.0670	0.8699
RTC2	20 mA	1.1794	0.9450
RTC3	20 mA	1.1763	1.0243
RTC4	20 mA	1.1545	0.8134
RTC5	20 mA	2.6024	1.9191
RTC6	20 mA	3.1834	2.4254
RTC7	15 mA	5.3388	3.4156
RTC8	10 mA	10.9367	7.3260

). For example, in RTC5, FID improves from 2.60 (noisy) to 1.92 (reconstructed), and in the most challenging RTC8 (10 mA), from 10.94 to 7.33, demonstrating improved alignment of reconstructed images with the high-dose feature distribution.

We also observed a similar trend in KID scores. Across RTC cases, both the mean and standard deviation of KID decrease after reconstruction (

Table 10. Kernel Inception Distance (KID) mean, standard deviation, and 95% confidence interval computed for each test case.

Test Case	Dose/ Tube Current	KID
		Noisy		Recon. (80 mA)
		Mean	Std. Dev.	Mean	Std. Dev.	95% CI (±)
RTC1	20 mA	1.0050	0.9504	0.9221	0.5529	0.0222
RTC2	20 mA	1.0852	0.9477	0.7501	0.4699	0.0480
RTC3	20 mA	1.1542	1.1247	0.9402	0.5423	0.0531
RTC4	20 mA	1.1293	1.0323	0.8289	0.4932	0.0729
RTC5	20 mA	4.6036	2.6247	1.7812	1.0626	0.0622
RTC6	20 mA	5.9372	3.5163	2.1289	1.1979	0.0702
RTC7	15 mA	10.1893	5.4318	4.0789	2.7578	0.1615
RTC8	10 mA	21.8039	8.9176	10.4164	5.4939	0.3218

), indicating higher perceptual similarity and reduced variability relative to 80 mA references.

NIQE scores, trained on domain-specific data and applied consistently across datasets, also decrease markedly after reconstruction (

Table 11. NIQE scores for noisy and reconstructed CT images across different test cases and 95% confidence interval. All values are reported with two decimal places.

Test Case	Noisy				Recon. (80 mA)
	Mean	Min	Max	Std. Dev.	Mean	Min	Max	Std. Dev.	95% CI (±)
RTC1	9.40	8.71	9.86	0.22	5.74	4.79	6.71	0.36	0.015
RTC2	9.52	9.08	9.91	0.16	6.17	5.36	7.01	0.32	0.033
RTC3	8.73	8.03	9.20	0.21	5.21	4.73	5.93	0.22	0.022
RTC4	9.19	8.58	9.59	0.20	5.37	4.81	6.33	0.32	0.047
RTC5	6.36	5.72	7.28	0.24	4.32	3.99	4.81	0.13	0.008
RTC6	6.76	5.90	7.83	0.33	4.85	4.11	5.98	0.32	0.019
RTC7	6.92	5.95	8.01	0.37	5.05	4.18	6.49	0.34	0.020
RTC8	7.06	6.05	7.74	0.25	4.60	4.01	5.30	0.22	0.013

), confirming improved perceptual naturalness and structural regularity.

Finally, Noise variance estimates (

Table 12. Estimated noise variance for noisy and reconstructed CT images across different test cases.

Test Case	Dose/ Tube Current	Noise Variance
		Noisy	Recon. (80 mA)
RTC1	20 mA	11.9385	8.8873
RTC2	20 mA	12.4323	8.7430
RTC3	20 mA	13.4390	9.0352
RTC4	20 mA	12.3812	8.9901
RTC5	20 mA	13.5279	8.5505
RTC6	20 mA	17.5892	8.7107
RTC7	15 mA	19.7846	9.3980
RTC8	10 mA	24.4319	11.4037

) show substantial reduction across RTC1–RTC8, with the most considerable improvements observed in lower-dose studies (e.g., RTC8: from 24.43 to 11.40), demonstrating stable noise suppression even under severe degradation.

5.4. Quantitative Image Quality Assessment

TEC-based analyses were performed on anatomically consistent locations (e.g., right ventricle, left superior pulmonary vein, descending aorta), and their respective Area Under the Curve (AUC) and correlation coefficients were obtained.

Table 13. AUC from Reconstructed Test Cases and respective Reference Test Cases for Left Superior Pulmonary Vein.

Test Case	Area Under the Curve (AUC)
	Reconstructed	Reference	Relative Error
RTC1	1478	1562	5.37%
RTC2	1658	1742	4.82%
RTC3	3868	3601	7.41%
RTC4	1296	1286	0.78%
RTC5	1660	1786	7.05%
RTC6	2287	2151	6.32%
RTC7	1824	2151	15.20%
RTC8	1381	2151	35.79%

and

Table 14. AUC from Reconstructed Test Cases and respective Reference Test Cases for Descending Aorta.

Test Case	Area Under the Curve (AUC)
	Reconstructed	Reference	Relative Error
RTC1	1537	1484	3.57%
RTC2	1955	2110	7.35%
RTC3	2715	2555	6.26%
RTC4	1278	1235	3.48%
RTC5	1874	1812	3.42%
RTC6	2285	2210	3.39%
RTC7	2394	2210	8.33%
RTC8	2518	2210	13.94%

report AUC values and relative errors for reconstructed TECs versus references. For all 20 mA cases (RTC1–RTC6), the relative AUC error remains below ∼10% in both regions, typically within 3–7%. Larger deviations are observed at 15 mA and 10 mA (RTC7–RTC8), reflecting the increased difficulty at ultra-low dose, but overall trends are preserved.

Correlation coefficients between reconstructed and reference TECs in the left superior pulmonary vein and descending aorta (

Table 15. Correlation Coefficient from Reconstructed Test Cases and respective Reference Test Cases for Left Superior Pulmonary Vein.

Test Case	Dose/Tube Current	Correlation Coefficient
RTC1	20 mA	0.9336
RTC2	20 mA	0.8807
RTC3	20 mA	0.7533
RTC4	20 mA	0.8275
RTC5	20 mA	0.9022
RTC6	20 mA	0.7947
RTC7	15 mA	0.7641
RTC8	10 mA	0.7429

and

Table 16. Correlation Coefficient from Reconstructed Test Cases and respective Reference Test Cases for Descending Aorta.

Test Case	Dose/Tube Current	Correlation Coefficient
RTC1	20 mA	0.9506
RTC2	20 mA	0.7342
RTC3	20 mA	0.7037
RTC4	20 mA	0.9386
RTC5	20 mA	0.8152
RTC6	20 mA	0.8033
RTC7	15 mA	0.6817
RTC8	10 mA	0.6033

) are generally high (often >0.8 and up to 0.95), especially at 20 mA. These findings confirm that the proposed method not only improves static image quality but also preserves temporal perfusion dynamics, which are critical for downstream quantitative analysis.

We have presented an example curve taken from the annotated location in Figure 2. In noisy, low-dose data, the bolus shape is often severely distorted, as shown in Figure 3, leading to unreliable perfusion parameters such as the Area Under the Curve (AUC) and Blood Flow Maps (BFM).

After unified correction, the reconstructed TECs regain a physiologically plausible bolus profile and align closely with the 80 mA as visible from reference (Figure 4).

This is clearly visible from Figure 5, when we superimpose one curve from the reference image and the other curve from the reconstructed image. Their closeness confirms that the reconstructed image can provide high-quality perfusion metrics.

5.5. Subjective Image Quality Assessment

Two domain experts (a medical biophysicist and a radiologist) independently assessed all reconstructed RTC datasets using the questionnaire in

Table 17. Summary of the subjective qualitative questionnaire designed for experts to assess reconstructed dynamic myocardial CTP images. The questionnaire comprises five evaluation categories (C1–C5) with seven questions (Q1–Q7), rated on a 5-point Likert scale, where scores range from 1 (poor or severe degradation) to 5 (excellent or no degradation). This questionnaire reflects a comprehensive scoring mechanism for the subjective assessment of image quality, anatomical fidelity, and clinical usability.

Cat#	Category	Q#	Question	Rating
C1	Overall Image Quality	Q1	How would you rate the overall image quality?	1 = Very poor 5 = Excellent
C2	Anatomical Detail and Sharpness	Q2	Are anatomical structures well delineated with respect to high-dose (80 mA) images?	1 = Not at all visible 5 = Very clear
C3	Diagnostic Confidence	Q3	Does the image provide sufficient detail to support diagnosis confidently?	1 = Not at all confident 5 = Very confident
		Q4	Would you consider this image diagnostically equivalent to a high-dose (80 mA) scan?	1 = Not at all 5 = Fully equivalent
C4	Artifacts and Distortions	Q5	Are there any visible artifacts introduced by the reconstruction process (e.g., unnatural textures, streaks)?	1 = Severe artifacts 5 = No artifacts
		Q6	Do you notice any distortion or deformation in anatomical regions?	1 = Severe distortion 5 = No distortion
C5	Temporal Consistency	Q7	Do the reconstructed time frames demonstrate realistic motion and changes in perfusion over time?	1 = Inconsistent 5 = Fully consistent

. Their reported mean scores are summarized in

Table 18. Summary of aggregated expert scores from the subjective qualitative questionnaire proposed to assess image quality and clinical relevance of reconstructed myocardial CTP images across all evaluated cases.

Q#	Assessor # 1	Assessor # 2	Mean Score
Q1	3.91	4.25	4.08
Q2	5.00	3.88	4.44
Q3	3.84	4.00	3.92
Q4	3.88	3.88	3.88
Q5	3.31	5.00	4.16
Q6	5.00	5.00	5.00
Q7	3.75	4.13	3.94

, and the overall distribution is shown in Figure 6.

Across all questions, mean scores are close to or above 4/5, indicating generally high perceived quality. Overall image quality (Q1) was rated at 4.08 on average, anatomical detail and sharpness (Q2) at 4.44, diagnostic confidence (Q3–Q4) around 3.9–4.0, and temporal consistency (Q7) at 3.94. The absence of noticeable distortions (Q6) was unanimously rated as 5.0 by both assessors. Slightly lower scores in some categories reflect residual streaks or texture artifacts in the most challenging ultra-low-dose cases. Still, both experts agreed that the reconstructed images are generally suitable for diagnostic interpretation in typical low-dose settings.

5.6. Visual Image Quality Assessment

Representative slices illustrating denoising performance are shown in Figure 7. Compared to the simulated 20 mA input, the reconstructed images exhibit markedly reduced noise while preserving myocardial and vascular structures, closely approximating the 80 mA reference. Figure 8 demonstrates the motion correction model’s ability to remove synthetic deformations and restore anatomical alignment. Figure 9 shows the unified model applied to real 20 mA data, where both noise and motion artifacts are substantially mitigated.

These visual examples align with the objective and subjective findings, illustrating that the unified framework can recover both anatomical detail and perfusion-relevant contrast patterns from low-dose myocardial CTP acquisitions.

6. Contributions

This work presents a unified deep-learning framework for low-dose myocardial CTP that jointly addresses residual motion artifacts and noise degradation. The framework integrates a previously validated 4D one-shot CNN–based motion correction model designed to capture nonrigid cardiac and respiratory motion across space and time, with a previously validated FFDNet-derived denoising model tailored to the mixed Poisson–Gaussian noise characteristics of low-dose CTP. A principled late-fusion ensemble strategy is introduced to combine the complementary strengths of the components, producing temporally coherent, noise-suppressed perfusion sequences suitable for quantitative analysis. The proposed approach is evaluated using a comprehensive image quality assessment protocol that includes reference-based, no-reference, and expert-based metrics, providing a rigorous, multi-perspective evaluation of unified noise and motion correction in dynamic myocardial CTP. To the best of our knowledge, this is among the first studies to investigate a learning-based unified framework for parallel ensemble denoising and motion correction in low-dose myocardial CTP using perfusion-preserving criteria within an image-domain integration setting under controlled porcine scans. While prior work, such as ASTRA4D [73] has addressed joint noise and motion compensation in clinical human datasets, these approaches relied on classical iterative reconstruction strategies and were typically evaluated against corresponding CCTA references. In contrast, the present work focuses on a deep learning–based image-domain integration framework and evaluates perfusion fidelity through temporal enhancement preservation and multi-perspective image quality assessment.

7. Limitations

This work has several limitations related to the experimental setup and evaluation that should be acknowledged. First, the proposed framework was evaluated on a limited number of porcine myocardial CTP studies acquired under controlled experimental conditions. Although this dataset enables careful analysis of noise–motion interactions and provides access to high-dose reference scans, it may not fully capture the variability encountered in real-life clinical practice, including differences in patient anatomy, heart rate variability, and acquisition protocols across scanners and institutions. Second, while both simulated (STC) and real low-dose (RTC) settings were considered, and the simulated noise closely resembled the true noise, as shown in experiments, the simulated noise model may not perfectly reflect all characteristics of true low-dose acquisitions, potentially affecting the observed performance under clinical noise conditions. In addition, although the proposed method is described as a unified processing framework, the motion correction component still relies on estimating deformation fields across time; in the studied data, these deformation vector fields exhibited limited variability, and an averaged deformation field was found to be sufficient in practice under the experimental setting for the studied dataset, but further investigation on more diverse datasets is required to validate the generality of this observation. Third, the experimental evaluation focused primarily on image-domain quality metrics, temporal consistency measures, and expert reader assessments; downstream clinical tasks such as perfusion quantification or diagnostic decision-making were not explicitly evaluated. Fourth, the ablation analysis in this study was limited to a comparison between serialized and unified processing strategies. While this comparison isolates the system-level impact of coordinated integration, more granular ablation studies, such as branch-specific evaluations under compound degradation or the explicit removal of the TEC-consistency constraint, would further clarify the contributions of individual components. These analyses were beyond the current experimental scope and dataset constraints but represent important directions for future investigation. In addition, the relationship between perceptual metrics and downstream diagnostic tasks in CT was not comprehensively evaluated, beyond the inclusion of expert visual scoring. Finally, comparisons were limited to representative serial pipelines rather than direct numerical comparisons with reconstruction-level joint motion–denoising methods, due to differences in problem formulation and data availability. While recent joint denoising–motion correction methods exist, many rely on task-specific reconstruction pipelines, lack perfusion-aware validation, or are not publicly reproducible; for these reasons, direct quantitative comparison was not feasible without introducing confounding assumptions. These factors suggest that the reported results should be interpreted as indicative of performance under the studied conditions. Further validation on larger, more diverse clinical datasets will be necessary to characterize the framework’s behaviour more fully.

8. Conclusions

We introduced a unified deep learning framework for parallel ensemble noise suppression and motion correction in low-dose dynamic myocardial CTP imaging. By training on synthetically degraded high-dose data and evaluating on both simulated and real low-dose acquisitions, the model demonstrated strong performance across reference-based and no-reference metrics, as well as expert visual assessment. The proposed approach effectively recovers structural detail, improves temporal consistency, and approximates high-dose image quality from substantially reduced radiation exposures. These findings highlight the method’s potential to support safer myocardial perfusion imaging while preserving image characteristics relevant to diagnostic interpretation. By enabling reliable noise and motion correction under reduced radiation dose, the proposed framework supports safer myocardial CTP protocols while preserving quantitative fidelity. This is particularly relevant for stress–rest studies and repeat imaging scenarios, where dose accumulation and motion artifacts remain critical barriers to routine clinical adoption.

9. Future Work

Besides addressing the limitations explicitly mentioned in Section 7, future research will focus on several directions. First, alternative network architectures and loss functions will be explored to enhance joint noise–motion correction further. Second, the framework will be extended to additional anatomical regions and low-dose clinical CT applications affected by physiological motion. Third, validation on human datasets will be pursued to assess clinical utility in real-world perfusion analysis. Finally, advanced ensemble strategies and broader comparisons with state-of-the-art sequential pipelines will be investigated to refine the integration of denoising and deformable registration. Although fixed ensemble weights were sufficient in this study, adaptive or patient-specific weighting strategies may further improve stability and will be explored in future work. Collectively, these efforts aim to support the development of a more broadly applicable and artifact-tolerant reconstruction pipeline for low-dose dynamic imaging.