High-resolution magnetic resonance angiography (MRA) data is clinically important and highly demanding for accurate diagnosis of various vascular disorders, such as stroke and aneurysm [1,2], and is also valuable in animal model studies. Despite its importance, it is highly challenging to acquire due to longer scanning time and prone to artifacts induced by motion and instrumental limitations [3,4]. In addition to longer scanning time, it also suffers from a low signal-to-noise ratio (SNR) proportional to the resolution or voxel size and inversely proportional to the field strength [5,6]. The presence of different types of noises prevented the method from visualizing fine details such as smaller vessels and bifurcations [5]. In addition, saturation effects during data acquisition, combined with noise, further degrade the overall quality of vascular information.

To increase the SNR and reduce the noise level in MRA data, different methods can be applied such as averaging and applying denoising algorithms [[7], [8], [9]]. Reducing the noise through averaging imposes a longer scanning time and cannot remove noise significantly with a small number of averaging. In the last few decades, many innovative methods have been developed to generate noise-free images from noisy images which can be generally categorized into spatial domain-based methods, transform-based methods, hybrid-based methods, and deep-learning-based methods [7]. Spatial domain-based filtering methods, such as Gaussian filtering, operate directly on the image domain and process each point using its local information. Compared to others, these methods are simple and fast but less effective at removing different types of noise present in the MRA data.

Transform-based filtering methods on the other hand operate on the different transformed domains such as the frequency domain and remove the noise based on the frequency distribution of the data. The most notable among these methods is denoising using wavelet transform, which represents the image at multiple scales [10,11]. Transform-based techniques are popular in medical image denoising due to their capacity for separating noise from structural information. The hybrid methods such as block-matching and 4D filtering (BM4D) methods operate in both spatial and transformed domains for effective reduction of noise from the 3D image [[12], [13], [14]]. Even if they are effective for removing different types of noise from higher-dimensional medical data, those methods are also known to have a higher computational cost [13]. Some classical denoising methods use the non-local means (NLM) algorithm, a seminal approach for image denoising that leverages the self-similarity of natural images by averaging similar patches across the entire image [15]. This approach effectively preserves fine details and structures, outperforming traditional local filtering methods [16]. However, it is computationally expensive due to exhaustive patch comparisons, struggles with heavy noise where patch similarity becomes unreliable, and is sensitive to parameter selection [16,17].

Currently, deep learning such as convolutional neural networks (CNN) based image denoising methods significantly improved the denoising process beyond the previous classical techniques [8,18,19]. Fully convolutional networks such as residual networks (ResNet), UNet, and others can be used for denoising tasks with different modes of training [20,21]. Depending on the presence or absence of ground truth or noise-free data, supervised, semi-supervised as well as self-supervised learning methods can be applied [8,9,18,21,22]. Supervised learning methods such as generative adversarial networks (GAN), diffusion models as well as a single feed-forward CNN model can be used [23,24]. However, it is always challenging to generate noise-free data experimentally and a processed image using the current state-of-the-art denoising method can be used as ground truth data [14,25]. On the other hand, self-supervised learning methods can be used without the need for noise-free data, but until now, their performance is lower with higher computational cost compared to the supervised methods [9,25]. In most research, supervised learning methods are commonly used. Many researchers have attempted to predict noise or noise-free images directly from the noisy images with different levels of efficiency [[15], [16], [17]]. Combining multiple denoising approaches or modalities has also shown promise. For instance, hybrid models that integrate wavelet denoising with CNN architectures can leverage the benefits of both approaches, preserving high-frequency details while effectively reducing noise [15].

CNN-based denoising methods have transformed medical image denoising by learning hierarchical features directly from data, unlike classical approaches like NLM and wavelet-based algorithms, which rely on hand-crafted features or local patch similarities. While classical methods are effective for certain tasks, they often struggle with adaptability and scalability to diverse imaging scenarios. Moreover, their computationally intensive operations such as patch similarity searches or wavelet decompositions become impractical for high-resolution or large-volume medical images, such as 3D MRA or computed tomography (CT) scans [10,15]. By contrast, deep learning models learn noise-reduction patterns tailored to specific imaging modalities from extensive datasets, enabling efficient inference on high-resolution images and significantly reducing computation time during deployment [26]. However, the computational demands of deep learning models vary with architecture, with lightweight models like UNet offering faster inference compared to deeper architectures like ResNet with larger residual blocks or Transformer-based denoisers, which require high-performance graphical processing units (GPUs), particularly during training.

Another key advantage of deep learning approaches is their ability to generalize across varying noise levels and imaging conditions, which classical methods often struggle with. Techniques like NLM or wavelet algorithms perform well under specific assumptions but may fail when dealing with structured or non-Gaussian noise, common in medical imaging [27]. Conversely, CNN-based models can adapt to these complexities when trained on representative datasets. Nonetheless, deep learning models also have limitations, such as requiring large, annotated datasets, substantial initial computational resources, and careful regularization to avoid overfitting. While classical and deep learning methods each have their merits, the latter's efficiency and adaptability make it a compelling choice for high-dimensional medical imaging tasks.

This study aims to develop effective denoising methods that address noise reduction challenges in MRA data while preserving critical anatomical fine details, even with a limited dataset from animal models. By leveraging advanced deep learning models, the focus is on achieving a balance between robust noise suppression and computational efficiency, ensuring practicality for real-world clinical applications with limited datasets. The proposed approach is designed to handle various noise types commonly encountered in 3D-MRA, such as Gaussian and Rician noise while minimizing artifacts and maintaining the fidelity of fine structures. Additionally, the computational efficiency of the methods ensures scalability for high-resolution 3D imaging datasets, paving the way for integration into time-sensitive preclinical research workflows.