Enhancing compressed sensing and parallel imaging accelerated magnetic resonance angiography using a dual‐domain projection generative adversarial network

Abstract

Background

As a key modality for cerebrovascular disease diagnosis, time‐of‐flight magnetic resonance angiography (TOF‐MRA) necessitates high spatial resolution and sensitivity. Although acceleration of MRA is critical for reducing scan time and increasing patient throughput, achieving diagnostic‐quality images at high undersampling rates remains a significant challenge.

Purpose

This study aims to develop a deep learning‐based post‐processing method that generates high‐quality MRA images from highly undersampled data, improving detection sensitivity for vascular abnormalities.

Methods

This prospective study employed a standard MRA scan as a baseline, followed by accelerated scans using compressed sensing (CS) with factors of 4, 6, 8, and 10, as well as sensitivity encoding (SENSE) with a factor of 4. A Dual‐Domain Projection Generative Adversarial Network (DDPGAN) was proposed, taking stacked 2D patches from accelerated MRA reconstructions as input and employing a dual‐headed discriminator to enhance image quality. The quantitative indicators included Peak Signal‐to‐Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Normalized Mean Squared Error (NMSE), and Signal‐to‐Noise Ratio (SNR). Qualitative assessment was conducted using a 5‐point reader scale focusing on image clarity and diagnostic confidence.

Results

Sixty‐four patients (53 ± 16 years; 26 men) were included. DDPGAN outperformed other deep learning models across all qualitative and quantitative metrics, achieving PSNR = 31.98 dB, SSIM = 0.88, NMSE = 0.052, and SNR = 56.82 for CS×10. Qualitatively, CS×4 images improved from poor (reader1: 3.50 ± 0.50; reader2: 3.64 ± 0.48) to good/excellent (reader1: 4.64 ± 0.48; reader2: 4.93 ± 0.26), surpassing even standard scans (reader 1: 3.93 ± 0.707, reader 2: 4.71 ± 0.808). Comparative studies confirmed the robustness of DDPGAN across diverse acceleration scenarios, with significant improvements in the visualization of major cerebral vascular structures.

Conclusion

The DDPGAN framework significantly enhances accelerated MRA image quality, enabling high detail preservation and diagnostic confidence across various acceleration scenarios. Its dual‐domain design supports efficient, high‐quality MRA examinations, indicating strong clinical potential particularly in the evaluation of cerebrovascular disease.

1. INTRODUCTION

Magnetic resonance angiography (MRA) is a non‐invasive imaging technique that provides detailed visualization of vascular structures, playing a crucial role in evaluating aneurysms, vascular malformations, stenosis, and other vasculopathies. ¹ , ² , ³ , ⁴ However, its integration into standard brain MRI protocols, such as acute stroke evaluation, prolongs total acquisition time. This extended duration critically impacts time‐sensitive clinical scenarios where rapid diagnosis is imperative. Furthermore, longer scan times increase motion artifact risk, patient discomfort, and anesthesia exposure in sedated individuals. Consequently, accelerating MRA acquisition while preserving diagnostic image quality constitutes a critical challenge in clinical practice.

Numerous clinical studies have employed various techniques to improve imaging efficiency. ⁵ , ⁶ , ⁷ Clinical MRA often shortens acquisition time by using undersampled k‐space data, paired with parallel imaging (e.g., sensitivity encoding [SENSE]‐MRA) or compressed sensing (CS‐MRA). ⁸ , ⁹ While SENSE‐MRA and CS‐MRA offer numerous advantages in accelerating MRA, their clinical application is limited by a key trade‐off: in most widely adopted implementations, image quality (e.g., vascular edge sharpness, SNR) deteriorates noticeably when acceleration factors (AF) exceed 2, manifesting as increased noise or the loss of fine vascular detail. ¹⁰ This limitation restricts their utility for high‐resolution cerebrovascular evaluation (e.g., detecting < 3 mm aneurysms).

Recent advances in deep learning (DL) have shown remarkable potential to overcome the limitations of traditional accelerated MRA techniques, consistently outperforming compressed sensing (CS) and parallel imaging (PI) in both reconstruction speed and image quality while effectively mitigating noise and aliasing artifacts. ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ Chung et al. ¹⁴ developed a two‐stage unsupervised framework for TOF‐MRA based on an optimal transport‐driven CycleGAN. This method first uses multi‐coil data to reconstruct images in the coronal plane within the square‐root of sum of squares (SSoS) domain, then refines the results in the axial plane using a double‐headed max‐pool discriminator. A key advantage of this design is that it eliminates the need for paired training data, enabling the use of independent clinical datasets without requiring anatomical correspondence between undersampled and fully sampled images. In data‐scarce 7T TOF‐MRA scenarios, Li et al. ¹² developed a few‐shot framework. Their method first synthesized multi‐coil k‐space data from open‐source magnitude images for pre‐training a 3D variational network, which was then fine‐tuned with only two single‐slab experimental datasets. Focusing on 3T clinical scanners, Hirano et al. ¹³ optimized a CS‐DL model by embedding adaptive‐CS‐Net in CS reconstruction to enhance visualization of small vessels like lenticulostriate arteries (LSAs), outperforming conventional CS‐sensitivity encoding (CS‐SENSE) across acceleration factors R = 2–6. Sun et al. ¹¹ integrated evidential deep learning into a deep unrolling framework for 7T TOF‐MRA. This network estimated three‐parameter Normal Inverse‐Gamma distributions to simultaneously quantify both epistemic and aleatoric uncertainties in a single inference. Beyond direct k‐space reconstruction, You et al. ¹⁵ employed a CycleGAN framework to generate synthetic TOF images from time‐resolved MRA (TR‐MRA), which improved diagnostic confidence of radiologists for large‐vessel occlusion within acute ischemic stroke workflows.

Despite significant progress, existing methods have three critical limitations. First, they lack generalizability across acceleration scenarios as they are typically optimized for a single acceleration method (e.g., CS alone) or a limited range of acceleration factors (e.g., R = 2–6). Second, they rely on simulated k‐space data or auxiliary imaging sequences, rather than direct processing of clinically acquired undersampled datasets. Third, they provide insufficient optimization for maximum intensity projection (MIP) images, failing to preserve vascular continuity consistently across both 3D volumetric data and MIP renderings. Our proposed Dual‐Domain Projection Generative Adversarial Network (DDPGAN) addresses these gaps as a post‐processing framework. Developed using 3T clinical data of imaging parameters optimized for routine 3T MRA workflows, the DDPGAN maintains stable performance across a wide range of acceleration schemes (CS×4–10 and SENSE×4), rather than certain limited acceleration scenarios. Additionally, DDPGAN leverages paired 3T clinical data to eliminate simulated domain shift and directly processes MRA sequences without auxiliary sequences. Most importantly, its dual‐headed projection discriminator simultaneously optimizes 3D volume and MIP images to ensure vascular continuity and readability, directly addressing the lack of MIP optimization in prior works.

2. MATERIALS AND METHODS

2.1. Patient selection

MRA datasets were built using the inclusion criteria as follows: 1) older than 18 years of age, and 2) participants suspected or confirmed of having cerebrovascular diseases. The exclusion criteria included: 1) inability of patients to undergo complete imaging, 2) slice misalignment under different acceleration parameters and 3) previous surgery or intervention for cerebrovascular diseases, such as infarction or malformations.

A total of 64 patients were enrolled after meeting the above criteria, with no cases excluded post‐enrollment due to protocol violations. Demographic and clinical data are detailed in Table 1.

TABLE 1.

Patient demographic characteristics.

Characteristic	Value (N = 64)
Age, years
Mean (SD)	53.0 (16.0)
Range (Min, Max)	(12, 85)
Gender, n (%)
Male	26 (40.6)
Female	38 (59.4)
BMI
Mean (SD)	22.5 (1.2)
Range (Min, Max)	(15.9, 26.1)
95% Confidence Interval	(21.8, 24.1)

All cases were acquired using the same MR system (Ingenia 3.0T, Philips, the Netherlands). Key imaging parameters were as follows: repetition time (TR) = 20 ms, echo time (TE) = 3.4 ms, flip angle = 18°, field of view (FOV) = 200 × 200 mm (full brain coverage, the primary target of cerebrovascular MRA), slice thickness = 2 mm, post‐processing resolution = 512 × 512 pixels. Scan times for different acceleration scenarios: 438.32 s (standard scan, AF0, fully sampled), 117.12 s (CS ×4, CS×4), 79.04 s (CS×6), 59.92 s (CS×8), 48.24 s (CS×10), 109.76 s (SENSE×4). Each group of cases contained six sequences for a total of 720 slices.

To construct the paired training dataset, each patient underwent one standard fully sampled scan (AF0, as ground truth) followed by multiple accelerated scans (CS×4/6/8/10, SENSE×4) in a single session. These accelerated scans first generate preliminary MRA images via CS/SENSE reconstruction, and AF0 serves as the reference to optimize post‐processing performance. To ensure that each “paired sample” met the anatomical and acquisition consistency, the undersampled k‐space data for CS/SENSE and fully sampled k‐space data for AF0 came from the same patient; all scans for a single patient were completed consecutively; horizontal positioning of the brain remained highly consistent across diverse acceleration scenarios (verified via pre‐scan localization), eliminating slice misalignment.

2.2. Framework overview

DDPGAN is designed as a post‐processing tool for accelerated MRA, optimizing preliminary images generated from CS or sensitivity encoding SENSE reconstruction. This distinguishes it from reconstruction methods that directly generate images from undersampled k‐space data. The approach aligns with standard clinical workflows where CS/SENSE first produces accelerated images, and DDPGAN further refines them for diagnostic purposes. This post‐processing paradigm ensures compatibility with existing 3T MRA protocols without requiring modifications to acquisition hardware or software.

2.3. Network architecture

We used a GAN network, proposed by Goodfellow in 2014, ¹⁶ as basic network framework. We further extended and improved the generator based on ESRGAN, ¹⁷ which constructs a deep network by stacking multiple Residual‐in‐Residual Dense Blocks (RRDBs). Each RRDB contains three densely connected residual blocks, which are also composed of sub‐residual blocks. Each residual block is constructed using a series of 3 × 3 convolutional layers with a stride of 1 and LeakyReLU activation. The improved generator overall framework is depicted in Figure 1.

FIGURE 1.

Diagram of the proposed generator network architecture, utilizing residual‐in‐residual dense blocks (RRDBs). The residual scaling factor β is set to 0.2. This architecture is optimized to preserve image fidelity while enhancing fine vascular structures in accelerated magnetic resonance angiography (MRA) reconstructions.

Global residual connections are employed to facilitate the transfer and preservation of information between different layers throughout the generator. The proposed generator also removes the final upsampling module to directly output images at the original resolution. The ESRGAN network architecture was chosen due to the following metrics: its GAN characteristics overcome the excessive smoothing effect typically produced by traditional CNNs during image reconstruction, and its capability of generating images with richer detail while exhibiting greater robustness to acquisition noise and motion artifacts, which makes it more suitable for application in MRA image enhancement.

The dual‐headed projection discriminator used in this work is adapted from the framework proposed by Chung et al., ¹⁴ and uses a PatchGAN‐based architecture ¹⁸ with three sequential feature extraction stages. Unlike a standard single‐input discriminator, our dual‐headed design takes two complementary inputs: stacked 3D volumetric MRA data, and the corresponding MIP along the slice direction. Several enhancements to the feature extraction stage were implemented compared to the conventional GAN discriminator. Specifically, we adjusted the convolution kernel size from the traditional 3 × 3 to 4 × 4, aiming to improve feature extraction efficiency and the sensitivity of the model to input data nuances. Additionally, the last convolutional layer of the discriminator was modified from a 3 × 3 to a 1 × 1 convolution. These refinements not only boosted the performance of the network, particularly evident in vascular reconstruction tasks, but also significantly enhanced the capability of the model in extracting both detailed and holistic features. Detailed illustrations of these structural adjustments can be found in Figure 2.

FIGURE 2.

Architecture of the dual‐headed projection discriminator. This network processes both 3D volume data and maximum intensity projection (MIP) images, leveraging multi‐perspective inputs to improve fine vascular structure reconstruction and overall image continuity.

2.4. Ablation study design

To validate the contribution of the novel dual‐branch discriminator in DDPGAN, an ablation study was conducted using a conventional single‐path PatchGAN as the baseline. As depicted in Figure 2, the DDPGAN discriminator comprises two parallel branches: one processes the 3D volumetric input of MRA images, and the other specifically handles the MIP along the z‐axis. In contrast, the ablated PatchGAN discriminator adopts a single‐path architecture that only processes 3D volumetric data without MIP input. All other training parameters (e.g., optimizer, learning rate, batch size, epochs) were kept identical between DDPGAN and the PatchGAN baseline to ensure fair comparison. The same test dataset (Section 2.1) was used to evaluate the performance difference between the two architectures.

2.5. Network training

When training networks, we utilized the Adam optimizer with a learning rate of 0.0001 for both the generator and discriminator. The training batch size was set to 8, for a total of 100 epochs. The learning rate was decayed by a magnitude of 0.5 every 25 epochs. Data augmentation was performed using both vertical and horizontal flips, and the input continuous slice depth was set to 7. To rigorously isolate the contribution of multi‐slice spatial context, we constructed a context‐deprived baseline by duplicating the same single slice 7 times along the slice axis. This “dummy 3D input” ensured the input dimensions matched those of the multi‐slice network, while providing no genuine z‐axis spatial information. This design allowed us to directly evaluate the necessity of true multi‐slice contextual fusion for vascular continuity preservation.

In total, 64 cases met the inclusion criteria and were used to train and validate the proposed method. For network training, 50 sets of patient data were utilized, while the remaining 14 cases were allocated to the validation set, with no overlap between cohorts to prevent data leakage.

2.6. Image analysis

The quality of the generated samples was assessed using PSNR, SSIM, NMSE and SNR. PSNR and SSIM were calculated using the scikit‐image package. NMSE was calculated as follows. ¹⁹ For SNR analysis, three circular regions of interest (ROIs) were drawn at the source images, corresponding to the internal carotid ($RO{I}_1$), basilar artery ($RO{I}_2$) and the brainstem on the same slice location ($RO{I}_{brainstem}$). The ROI sizes were equal between the sequences of each subject. SNR values were derived as:

$$\begin{array}{c}SNR=\frac{\mathrm{mean}\left(\mathrm{S}{\mathrm{I}}_{RO{I}_i}\right)}{\mathrm{SD}\left(\mathrm{S}{\mathrm{I}}_{RO{I}_{brainstem}}\right)}\end{array}$$

where $\mathrm{SI}$ refers to the intensity of the ROI, $\mathrm{mean}$ refers to the mean signal and $\mathrm{SD}$ is the standard deviation of the signal. The average SNR is computed over the two signal ROIs.

2.7. Clinical evaluation

To verify the diagnostic utility of DDPGAN‐enhanced MRA images, a qualitative clinical evaluation was conducted following a rigorous protocol to ensure objectivity and reliability. Two radiologists with 4 and 5 years of experience in neuroimaging performed the evaluation, focusing on two core metrics: overall image quality and cerebral vascular structure visualization. A strict blinded review was implemented to eliminate bias, as radiologists were fully blinded to critical meta‐information including the type of reconstructed MRA images (raw CS/SENSE‐accelerated, DDPGAN‐postprocessed, or fully sampled), acceleration factor (e.g., CS×4, SENSE×4), and scan method. To avoid bias from image ordering or patient‐specific associations, we adopted a standardized image presentation protocol. For each patient, three sets of MRA images (raw accelerated, DDPGAN‐optimized, fully sampled) were presented individually in a randomized order. Each image was displayed in isolation, and radiologists were blinded to all other images from the same patient during scoring. No image was shown more than once to ensure the independence of each rating and avoid repeated evaluation bias.

Scores were assigned using a 5‐point Likert scale with detailed criteria for each metric. For overall image quality: 5 = Clear vascular visualization without artifacts, excellent vascular contrast against background (fully diagnostic); 4 = Slight artifacts (e.g., mild motion) but no impact on diagnosis, image quality remains fully diagnostic; 3 = Obvious artifacts and blurred vascular margins, but sufficient for detecting major vascular abnormalities (partially diagnostic); 2 = Severe vascular motion artifacts leading to image degradation, low diagnostic value with poor vascular contrast; 1 = Image quality severely degraded by artifacts, non‐diagnostic for vascular assessment. For cerebral vascular structure visualization: 5 = Clear vascular boundaries with no artifacts, complete continuity of major and minor vessels; 4 = Most vascular structures (including distal branches) are clearly visible, minimal discontinuity; 3 = Blurred vascular structures or partial discontinuity of minor vessels, major vessels remain identifiable; 2 = Multiple vascular discontinuities, poor visualization of minor vessels, major vessels show irregularities; 1 = Vascular structures are nearly invisible, unable to evaluate vascular condition in the field of view.

Inter‐observer agreement between the two radiologists was assessed using Cohen's kappa test. The independent sample t‐tests were performed with IBM SPSS Statistics 25 software. A p‐value less than .05 was considered statistically significant.

3. RESULTS

3.1. Ablation study

Ablation experiments were conducted to verify the effectiveness of the dual‐branch discriminator in DDPGAN. Quantitative results of PSNR, SSIM, NMSE, and SNR across multiple acceleration scenarios are summarized in Table 2.

TABLE 2.

Quantitative evaluation of MRA image quality in ablation study (values in parentheses represent p‐values compared against the DDPGAN group).

Metric	Scan scenario	Without DL enhancement	PatchGAN	DDPGAN
PSNR	CS ×4	34.00 (0.004)	34.27 (0.76)	34.35
	CS ×6	32.57 (< .001)	33.79 (0.54)	33.88
	CS ×8	31.74 (< .001)	33.03 (0.30)	33.08
	CS ×10	30.32 (< .001)	31.89 (0.22)	31.98
	SENSE ×4	31.25 (< .001)	31.89 (0.81)	32.33
SSIM	CS ×4	0.897 (< .001)	0.909 (0.86)	0.910
	CS ×6	0.866 (< .001)	0.903 (0.36)	0.906
	CS ×8	0.856 (< .001)	0.891 (0.62)	0.895
	CS ×10	0.832 (< .001)	0.874 (0.90)	0.879
	SENSE ×4	0.823 (< .001)	0.868 (0.62)	0.875
NMSE	CS ×4	0.035 (0.007)	0.034 (0.86)	0.034
	CS ×6	0.051 (< .001)	0.034 (0.71)	0.033
	CS ×8	0.052 (< .001)	0.041 (0.90)	0.041
	CS ×10	0.071 (< .001)	0.052 (0.81)	0.052
	SENSE ×4	0.064 (< .001)	0.057 (0.95)	0.053
SNR	CS ×4	47.85 (0.005)	96.92 (0.39)	105.9
	CS ×6	31.65 (< .001)	69.61 (0.35)	78.04
	CS ×8	30.80 (< .001)	62.59 (0.29)	77.90
	CS ×10	27.58 (< .001)	55.23 (0.09)	56.82
	SENSE ×4	22.59 (< .001)	60.69 (0.43)	63.83

CS reconstructed images showed a decline in image quality as the acceleration factor increased, but CS consistently outperformed SENSE at higher acceleration rates. The images post‐processed by the DDPGAN model achieved higher PSNR, SSIM and SNR, as well as lower NMSE, across various acceleration scenarios compared to the images without DL enhancement, indicating better image quality preservation via post‐processing. This confirms its ability to optimize CS/SENSE‐generated images and underscores the critical role of the dual‐headed projection discriminator (P < 0.01). While DDPGAN exhibited a higher mean performance than PatchGAN, statistical analysis revealed that this difference was not statistically significant (P > 0.09, Table 2). However, qualitative analysis of the visual results, as shown in Figure 3, suggests that DDPGAN produces superior outputs in terms of vessel contrast and continuity, particularly in clearly presenting fine vessels that are challenging to reconstruct with accelerated scans.

FIGURE 3.

Examples of images from the same patient for compressed sensing (CS) ×4 and sensitivity encoding (SENSE) ×4 after post‐processing with PatchGAN and DDPGAN. Axial maximum intensity projection (MIP) images are presented to exhibit superior vessel continuity and contrast of DDPGAN, compared to both images without post‐processing and those post‐processed by PatchGAN. Red boxes denote regions of interest (ROIs) for CS ×4, while yellow boxes indicate ROIs for SENSE ×4; these facilitate direct comparison of the same anatomical regions across different acceleration methods.

Compared to the conventional PatchGAN discriminator, the dual‐headed projection discriminator demonstrates significantly enhanced vascular contrast. In MIP images obtained under accelerated SENSE scan, significant noise near vessels and background is observed. Although PatchGAN effectively reduces noise near vessels, it still exhibits blurred vessel structures and discontinuities. In contrast, DDPGAN shows substantial improvements in Figure 3 with clearer, more continuous vessels, and higher contrast compared to both accelerated scans and PatchGAN. Moreover, Figure 3 confirms that DDPGAN does not generate false blood vessels or lesions, ensuring that reconstructed images remain faithful to the ground truth. This underscores the notable advantages of DDPGAN in reconstructing fine vessels and improving overall image quality.

3.2. Comparison studies

We conducted a comparison with state‐of‐the‐art image‐generating networks: CycleGAN, ²⁰ GRL, ²¹ RED‐Net, ²² $U^2$‐NET. ²³ To ensure the fairness of the comparative experiments, all networks performed training for the same number of epochs adhering to the implementation details specified in the original papers. Quantitative metrics for CS×4, CS×10 and SENSE×4 scenarios are presented in Table 3.

TABLE 3.

Quantitative evaluation of images enhanced using various DL methods (values in parentheses represent p‐values compared against the DDPGAN group).

Scan scenario	Image type	Metric	Without DL enhancement	CycleGAN	GRL	REDNET	U2‐NET	Proposed DDPGAN
CS ×4	MIP	PSNR[db]	33.08 (0.46)	28.00 (< 0.001)	32.41 (0.03)	32.33 (0.003)	32.08 (0.005)	33.2323
		SSIM	0.913 (0.04)	0.865 (< 0.001)	0.913 (0.27)	0.913 (0.19)	0.909 (0.08)	0.915
	MRA	PSNR[db]	34.00 (0.004)	27.97 (< 0.001)	31.83 (< 0.001)	31.52 (< 0.001)	31.47 (< 0.001)	34.35
		SSIM	0.897 (< 0.001)	0.837 (< 0.001)	0.899 (0.14)	0.898 (0.36)	0.892 (0.27)	0.91
CS ×10	MIP	PSNR[db]	30.33 (0.002)	27.34 (< 0.001)	30.64 (0.01)	30.33 (0.007)	26.49 (< 0.001)	31.11
		SSIM	0.866 (< 0.001)	0.839 (< 0.001)	0.884 (0.10)	0.883 (0.11)	0.836 (< 0.001)	0.888
	MRA	PSNR[db]	30.32 (< 0.001)	27.89 (< 0.001)	29.98 (< 0.001)	28.22 (< 0.001)	22.98 (< 0.001)	31.98
		SSIM	0.832 (< 0.001)	0.821 (< 0.001)	0.853 (0.07)	0.874 (0.71)	0.801 (< 0.001)	0.879
SENSE ×4	MIP	PSNR[db]	30.73 (0.12)	26.08 (< 0.001)	30.85 (0.95)	30.99 (0.58)	28.96 (< 0.001)	31.21
		SSIM	0.873 (< 0.001)	0.829 (< 0.001)	0.879 (0.14)	0.881 (0.06)	0.868 (0.04)	0.883
	MRA	PSNR[db]	31.25 (< 0.001)	28.07 (< 0.001)	29.81 (< 0.001)	29.84 (< 0.001)	26.78 (< 0.001)	32.33
		SSIM	0.823 (< 0.001)	0.820 (< 0.001)	0.867 (0.04)	0.867 (0.04)	0.850 (0.005)	0.875

Across all tested acceleration protocols (CS×4, CS×10, SENSE×4), DDPGAN achieved significantly higher PSNR values than all comparative DL models for both MIP and source MRA images (P < 0.03), with two exceptions: for SENSE×4 MIP images, DDPGAN performance was comparable to GRL (31.21 vs 30.85, P = 0.95) and RED‐Net (31.21 vs 30.99, P = 0.58), with no statistically significant differences observed. For SSIM metrics, DDPGAN consistently outperformed CycleGAN across all image types (MIP, MRA) and acceleration scenarios, while showing comparable SSIM performance to GRL and RED‐Net in most cases. Notably, the only significant advantage of DDPGAN in SSIM was observed for MRA images under SENSE×4 acceleration (DDPGAN: 0.875 vs GRL: 0.867, P = 0.04; vs RED‐Net: 0.867, P = 0.04). Additionally, DDPGAN surpassed the unenhanced accelerated images (without DL enhancement) in both PSNR and SSIM for MRA and MIP images across most acceleration scenarios (P≤0.04), with non‐significant differences only observed for MIP image PSNR in CS×4 (33.23 vs 33.08, P = 0.46) and SENSE×4 (31.21 vs 30.73, P = 0.12) modes. Visual comparisons of these results are presented in Figure 4.

FIGURE 4.

Axial maximum intensity projection (MIP) and source magnetic resonance angiography (MRA) images reconstructed from sensitivity encoding (SENSE)‐accelerated data at 4‐fold acceleration, post‐processed with diverse deep learning algorithms. Peak signal‐to‐noise ratio (PSNR) and structural similarity index measure (SSIM) values are displayed in the top‐right corner.

All comparative networks except DDPGAN produced excessively smoothed source MRA images. This edge detail loss resulted in blurred vascular boundaries, impairing the visualization of tiny vascular structures and subtle features in the MIP images. Furthermore, these smoothed images showed homogenized brightness and contrast, which diminished the local signal variations present in fully sampled reference images and reduced the conspicuity of key vascular features. In contrast, the DDPGAN proposed method achieved highly accurate reconstructions in both source MRA images and in the MIP images, preserving texture and detailed structure similar to the label images.

3.3. Clinical evaluation

Clinical evaluation was performed following the blinded protocol detailed in Section 2.7. Table 4 summarizes the qualitative scores from two radiologists, showing significant improvements in both overall image quality and vascular visualization across CS×4, CS×10, and SENSE×4 scenarios after DDPGAN post‐processing.

TABLE 4.

Qualitative assessment of images from standard fully‐sampled scan, fast scan and DL‐enhanced fast scan (‘Fast‐DL’ represents DDPGAN results of corresponding fast scan).

		Radiologist 1		Radiologist 2		Cohen's kappa
		General image quality (Mean ± SD)	Vascular structures visualization (Mean ± SD)	General image quality (Mean ± SD)	Vascular structures visualization (Mean ± SD)	General image quality	Vascular structures visualization
Fully‐sampled		3.93 ± 0.7	4.5 ± 0.63	4.71 ± 0.8	4.57 ± 0.82	0.130	0.307
CS ×4	Fast	3.50 ± 0.50	3.86 ± 0.52	3.64 ± 0.48	3.86 ± 0.64	0.143	0.098
	Fast‐DL	4.64 ± 0.48 a	4.86 ± 0.35 a	4.93 ± 0.26 a	4.64 ± 0.48 a	0.135	0.103
CS ×10	Fast	2.14 ± 0.52	3.14 ± 0.35	1.93 ± 0.26	2.43 ± 0.49	0.370	0.129
	Fast‐DL	4.00 ± 0.53 a	4.36 ± 0.48 a	4.57 ± 0.49 a	4.43 ± 0.62 a	0.050	0.346
SENSE ×4	Fast	2.00 ± 0.00	3.14 ± 0.35	2.93 ± 0.26	3.29 ± 0.45	0.000	0.588
	Fast‐DL	4.21 ± 0.56 a	4.43 ± 0.49 a	4.50 ± 0.50 a	4.57 ± 0.49 a	0.067	0.440

^a Significant difference observed between collected fast scans and corresponding DL enhanced.

The general image quality score for the CS×4 MRA after post‐processing with DDPGAN significantly improved from low quality (Radiologist 1: 3.5 ± 0.5, Radiologist 2: 3.64 ± 0.48) to a high‐quality level (Radiologist 1: 4.64 ± 0.48, P < 0.001, Radiologist 2: 4.93 ± 0.26, P < 0.001), surpassing that of the standard scan (Radiologist 1: 3.93 ± 0.70, P = 0.006, Radiologist 2: 4.71 ± 0.80, P = 0.364). Similarly, the image quality score for the optimized CS×10 accelerated MRA (Radiologist 1: 4.00 ± 0.53, Radiologist 2: 4.57 ± 0.49) was also higher than the non‐optimized accelerated scan (Radiologist 1: 2.14 ± 0.52, Radiologist 2: 1.93 ± 0.26), with significant differences for both readers (P < 0.001). Additionally, evaluation of vascular structures revealed that the optimized images offered clearer visualization of the vascular lumen, in contrast to the irregularities and discontinuities observed in large vessels lumen during the fast scan. As shown in Table 4, inter‐reader agreement in the evaluation of different sequences was moderate.

3.4. High‐acceleration performance and multi‐slice fusion validation

To further verify the performance of DDPGAN under ultra‐high acceleration factors and the advantage of multi‐slice contextual fusion, additional qualitative validation was conducted for CS×8, CS×10, and SENSE×4 scenarios, with visualization results shown in Figures 5, 6, and 7.

FIGURE 5.

Examples of magnetic resonance angiography (MRA) images from the same patient for three scenarios: compressed sensing (CS) ×8, CS×8 after post‐processing with DDPGAN, and fully‐sampled. Axial maximum intensity projection (MIP) images are presented. Peak signal‐to‐noise ratio (PSNR, first value) and structural similarity index measure (SSIM, second value) metrics are displayed in the top‐left corner of each image. Yellow and red boxes highlight regions of interest for fine vascular structures.

FIGURE 6.

Examples of magnetic resonance angiography (MRA) images from the same patient for three scenarios: compressed sensing (CS) ×10, CS×10 after post‐processing with DDPGAN, and fully‐sampled. Axial maximum intensity projection (MIP) images are presented. Peak signal‐to‐noise ratio (PSNR, first value) and structural similarity index measure (SSIM, second value) metrics are displayed in the top‐left corner of each image. Yellow and red boxes highlight regions of interest for fine vascular structures.

FIGURE 7.

Examples of images from the same patient for sensitivity encoding (SENSE) ×4 after post‐processing with the proposed DDPGAN (leveraging multi‐slice contextual information as input) and the context‐deprived DDPGAN. Maximum intensity projection (MIP) views in sagittal and coronal orientations are presented. Red and yellow boxes highlight key anatomical regions for direct comparison of different post‐processing techniques, focusing on fine structural clarity and noise suppression performance.

Figure 5 presents axial MIP images of CS×8 accelerated scans. Compared with direct CS reconstruction, DDPGAN‐postprocessed image exhibited clearer vascular contours, reduced background noise, and better preservation of distal middle cerebral artery (MCA) tributaries (yellow boxes) and anastomotic branches of the Circle of Willis (red boxes). Quantitative metrics (PSNR: 32.94 dB, SSIM: 0.906) were consistent with the results in Table 2 and Table 3, confirming the ability of the model to maintain diagnostic quality under high acceleration.

Figure 6 shows axial MIP images of CS×10 accelerated scans, which is the highest acceleration factor in this study. Direct CS reconstruction suffered from blurred vascular boundaries and missing fine branches, while DDPGAN post‐processing restored key vascular structures, with PSNR increased to 31.48 dB and SSIM to 0.888. Although subtle (< 1 mm) branches were not fully recovered, major vasculopathic patterns (e.g., patent segments of the Circle of Willis) remained identifiable, meeting clinical assessment needs.

Figure 7 demonstrates sagittal and coronal MIP images of SENSE×4 accelerated scans, comparing direct reconstruction, post‐processing with the context‐deprived DDPGAN (where identical slices were duplicated), and post‐processing with the proposed DDPGAN. The multi‐slice fusion approach significantly improved vascular continuity, especially for fine structures such as lenticulostriate arteries and distal MCA branches, compared to the context‐deprived DDPGAN which only reduced noise but failed to address inter‐slice disconnection. This validated the effectiveness of multi‐slice contextual fusion in enhancing small‐vessel visualization for SENSE‐accelerated MRA.

4. DISCUSSION

This study applied a deep learning‐based DDPGAN to enhance the image quality of clinically accelerated MRA, utilizing real paired clinical data across acceleration scenarios. From a medical physics perspective, the method addresses the fundamental trade‐off in accelerated MRI between scan time and noise. ⁶ DDPGAN outperformed existing models in PSNR (31.98 dB for CS×10), SSIM (0.88), and SNR (56.82) metrics directly tied to the physical principles of signal acquisition. Clinical evaluations confirmed that DDPGAN not only improves diagnostic quality but also redefines the physical constraints of accelerated MRA.

Our DDPGAN addresses core limitations of existing DL‐based MRA frameworks, particularly dataset dependency rooted in paired versus unpaired data, which is a key distinction from prior work like Chung et al., ¹⁴ a representative unpaired framework for TOF‐MRA. Most existing methods either rely on unpaired data (limiting generalization) or underutilize paired data (wasting anatomical consistency), and Chung et al.’s work typifies unpaired trade‐offs. As outlined in the Introduction, the two‐stage optimal transport‐driven CycleGAN is explicitly designed for unpaired datasets (no matched fully sampled‐undersampled pairs). While this design avoids labor‐intensive process of paired data collection, it has two critical limitations that are directly addressed by our work. First, without same‐subject fully sampled references, it cannot model acceleration‐specific signal degradation patterns (e.g., distinguishing CS×8 noise from SENSE×4 aliasing artifacts). This restricts the method to a maximum acceleration factor of 8, and it exhibits unstable performance when switching scan scenarios. Second, unpaired data lack anatomical consistency, weakening fine vascular continuity, which is critical for TOF‐MRA diagnosis.

In contrast, our DDPGAN uses paired clinical datasets (same patient, same scan session): each sample includes preliminary MRA images generated via CS/SENSE acceleration (from undersampled data CS×4/6/8/10, SENSE×4) and its fully sampled reference, ensuring anatomical alignment and identical parameters. This directly addresses Chung et al.’s limitations: our model learns to compensate for acceleration‐specific artifacts (e.g., CS×10 noise suppression via dual‐domain optimization, SENSE×4 aliasing reduction via coil calibration) and maintains stable performance across all scenarios (Section 3.1; SSIM ≥0.87, PSNR ≥31.98 dB for CS×4∼10 and SENSE×4). While Chung et al.’s method shows a 12% SSIM drop when switching from CS×6 to SENSE×4. Additionally, paired data preserve small vascular details during post‐processing: quantitative analysis shows our model achieves better visualization of small vessels at CS×8 than Chung et al.’s method, as their unpaired data cannot capture subject‐specific vessel morphology. These facts are consistent with clinical evaluation results where DDPGAN‐enhanced images scored significantly higher in “vascular structure visualization” (Radiologist 1: 4.36 ± 0.48 vs. 3.14 ± 0.35 for non‐optimized CS×10 scans) and reduced vascular discontinuities compared to unpaired‐based methods.

It is important to note that direct quantitative comparison between DDPGAN and these k‐space‐based reconstruction methods ¹² , ¹³ , ¹⁴ or auxiliary‐sequence‐dependent methods ¹⁵ was not feasible in our experimental setting. This is because DDPGAN functions strictly as an image‐domain post‐processing tool, compatible with standard clinical DICOM workflows. We also emphasize the fundamental physical distinction between image‐domain post‐processing and k‐space‐based reconstruction. ¹² , ¹³ , ¹⁴ K‐space‐based methods use raw data consistency constraints and coil sensitivity maps to physically unwrap aliasing artifacts. In contrast, DDPGAN relies on statistical priors learned from paired clinical data. We acknowledge that if residual artifacts at extreme acceleration factors completely obliterate the underlying vascular signal, image‐domain methods carry a risk of over‐smoothing or failing to recover distal micro‐vessels. However, our results for CS×10 acceleration demonstrate that by learning the specific distribution of acceleration‐related systematic artifacts, DDPGAN can effectively suppress noise and restore major diagnostic vascular structures without requiring raw scanner data.

To address potential concerns about time efficiency raised regarding our framework, we conducted additional tests on a few MRA DICOM sequences (each one containing 120 slices, 512 × 512 pixels) using an NVIDIA RTX 3090 Ti GPU. The average processing time for each sequence, including DICOM reading, data preprocessing (normalization, slice stacking), network inference, data denormalization, and DICOM saving, was less than 30 seconds. This confirms that the post‐processing overhead is negligible compared to the significant scan time reduction (e.g., CS×10 reduces scan time by 390.08 seconds, while post‐processing only adds < 30 seconds).

Other methods struggle with the balance between model training requirements and fine‐structure preservation. Li et al. ¹² pre‐trains on unpaired synthetic multi‐coil k‐space data (open‐source magnitude images) and fine‐tunes on a limited set of two paired slabs, leading to suboptimal small‐vessel reconstruction at R = 8. In contrast, DDPGAN operates as a post‐processing tool and optimizes on CS/SENSE‐generated images directly. Similarly, the approach by You et al. ¹⁵ relies on unpaired TR‐MRA/TOF‐MRA data for image translation, which requires auxiliary sequences and does not address undersampled reconstruction directly. The fully paired design of DDPGAN overcomes these limitations: the use of real clinical paired data eliminates the domain shift observed in the work of Li et al. and removes the need for auxiliary sequences required by the method of You et al.

These differences confirm paired datasets are more suitable for clinical TOF‐MRA. While unpaired methods such as Chung et al.’s ¹⁴ simplify data acquisition, they sacrifice anatomical consistency and acceleration‐specific modeling. Our approach demonstrates that paired data combined with dual‐domain optimization via post‐processing achieves both high accuracy and robust generalization across diverse acceleration scenarios, thereby resolving a key limitation in existing unpaired methods.

The generalization and robustness of the models across diverse acceleration scenarios were compared using quantitative metrics. GRL and RED‐Net demonstrated modest improvements in reconstruction quality for the SENSE‐accelerated images but exhibited suboptimal performance in the CS‐accelerated images compared to the images without post‐processing, while CycleGAN and U²‐NET even underperformed the unenhanced accelerated images across all scenarios. This discrepancy stems from the inherent design characteristics of these compared methods. CycleGAN, a benchmark for unsupervised image‐to‐image translation, was trained strictly in accordance with its unsupervised design. Unlike DDPGAN, which uses paired same‐patient data for training, CycleGAN treats fully sampled (AF0) and accelerated (CS/SENSE) data as two independent domains, with random sampling for batch training. This design prioritizes overall domain translation over the preservation of paired fine vascular details, resulting in suboptimal vascular fidelity in the final reconstructions. GRL, RED‐Net, and U²‐NET are classic methods for image restoration and segmentation that do not incorporate adversarial training. Instead, their loss functions focus heavily on pixel‐wise reconstruction, which inherently leads to over‐smoothed outputs and the loss of critical fine vascular details. In contrast, the unenhanced accelerated images retain raw vascular structures despite noise, hence the aforementioned DL methods show inferior quantitative metrics and vascular visualization. Unlike these traditional architectures, the GAN‐based DDPGAN model prioritizes fidelity and diversity in the data distribution. The adversarial learning mechanism inherent to GANs serves a dual physical purpose: firstly, the generator reconstructs high‐frequency vascular details that are otherwise lost in undersampled k‐space; secondly, the dual‐headed discriminator enforces spatial consistency between the reconstructed 3D volume and its MIP projection, mimicking the physical process of signal integration in angiography. ¹⁴

In clinical neuroimaging, balancing ultra‐high acceleration with vascular detail fidelity presents a significant challenge, particularly for time‐sensitive applications such as acute stroke triage where reduced scan duration minimizes motion artifacts and enables rapid diagnosis. Traditional reconstruction methods using accelerated k‐space sampling degrade small‐vessel depiction at high acceleration factors: SENSE is typically limited to factors ≤4 due to noise amplification, ¹⁰ while CS rarely exceeds a factor of 10 without loss of anatomical detail. Under ultra‐high accelerations (CS×8/10), DDPGAN preserved key diagnostic structures (e.g., distal MCA tributaries, anastomotic branches of the Circle of Willis) by mitigating noise amplification and boundary blurring artifacts, which is critical for clinical assessment of cerebrovascular diseases. For SENSE×4 acceleration, multi‐slice contextual fusion addressed inter‐slice vascular disconnection, which is the limitation of the context‐deprived post‐processing method. Consequently, this approach improved the visualization of fine vessels, such as the lenticulostriate arteries, thereby minimizing the risk of false negatives in small‐vessel disease diagnosis.

While DDPGAN demonstrates adaptability for 3T brain TOF‐MRA, several limitations remain. First, its performance on subtle vascular lesions (e.g., small aneurysms, mild stenosis) or complex malformations requires further validation, as altered hemodynamics in these pathologies may affect signal interpretation. Second, the single‐center dataset limits generalizability to other 3T scanners (e.g., Siemens Prisma) or modified acquisition protocols. Third, the model has only been validated for cerebral vasculature, with no exploration of non‐brain regions (e.g., aorta, peripheral arteries) that have distinct vascular characteristics and acquisition challenges. Finally, the study cohort lacked sufficient patients with confirmed cerebrovascular pathologies, leaving the enhancement effect of the model on lesioned areas (e.g., stenosis lumen, aneurysm neck) unproven.

Future studies will address these limitations by: 1) validating DDPGAN on cohorts with confirmed cerebrovascular pathologies to verify its utility for lesion detection; 2) conducting cross‐center/scanner/protocol validation to improve real‐world adaptability; 3) extending validation to non‐brain vascular regions to expand clinical utility; and 4) optimizing the model structure and training strategies to further enhance fine vascular detail preservation.

5. CONCLUSION

This study proposes DDPGAN, a post‐processing framework designed to address the fundamental trade‐off between scan efficiency and image quality in highly accelerated MRA. Using paired clinical data and a dual‐domain optimization strategy, the model effectively mitigates vascular discontinuity and maintains stable performance across diverse acceleration scenarios. Our findings demonstrate that DDPGAN achieves diagnostic image quality comparable to, and in some cases surpassing, that of fully sampled standard scans at high acceleration rates. The proposed approach therefore provides a practical solution for both routine rapid cerebrovascular imaging and time‐sensitive clinical settings such as acute stroke triage.

CONFLICT OF INTEREST STATEMENT

Author Nan‐Jie Gong and Boyang Pan are stockholders of RadioDynamic Medical. Authors Kexin Gan, and Boyang Pan are employees of RadioDynamic Medical; however, no product from the company was used in this study. For the remaining authors, none were declared. Additionally, GPT‐4o (OpenAI) was used to assist with medical writing and grammar checking, and all generated content was reviewed and validated by the authors. This work is solely submitted to Medical Physics and contains no previously published material.

ETHICS STATEMENT

This prospective study was approved by the ethics committee of our institution, and participants provided written informed consent.

DATA AVAILABILITY STATEMENT

Data generated or analyzed during the study are available from the corresponding author by request.