Image restoration driven by dual-scale prior.

Iterative learning for joint image denoising and motion artifact correction of 3D brain MRI.

TiDE-Net: A time-guided dual-encoder ResUNet for Positron Emission Tomography (PET) image denoising.

Medical image denoising is crucial for enhancing the diagnostic accuracy of CT and MRI images. This paper presents a modular hybrid framework that combines multiscale decomposition techniques (Empirical Mode Decomposition, Variational Mode Decomposition, Bidimensional EMD, and Multivariate EMD) with curvelet transform thresholding and traditional spatial filters. The methodology was assessed using a phantom dataset containing regulated Rician noise, clinical CT images rebuilt with sharp (B50f) and medium (B46f) kernels, and MRI scans obtained at various GRAPPA acceleration factors. In phantom trials, MEMD–Curvelet attained the highest SSIM (0.964) and PSNR (28.35 dB), while preserving commendable perceptual scores (NIQE approximately 7.55, BRISQUE around 38.8). In CT images, VMD–Curvelet and MEMD–Curvelet consistently outperformed classical filters, achieving SSIM values over 0.95 and PSNR values above 28 dB, even with sharp-kernel reconstructions. In MRI datasets, MEMD–Curvelet and BEMD–Curvelet reduced perceptual distortion, decreasing NIQE by up to 15% and BRISQUE by 20% compared to Gaussian and median filtering. Deep learning baselines validated the framework’s competitiveness: BM3D attained high fidelity but necessitated 6.65 s per slice, while DnCNN delivered equivalent SSIM (0.958) with a diminished runtime of 2.33 s. The results indicate that the proposed framework excels at noise reduction and structure preservation across various imaging settings, surpassing independent filtering and transform-only methods. Its versatility and efficiency underscore its potential for therapeutic integration in situations necessitating high-quality denoising under limited acquisition conditions.

Hydroxyapatite (HAP) is widely utilized in various applications, where its properties are strongly regulated by ionic substitution. However, the atomic-scale structural origins of such modulation remain poorly understood. Although high-resolution transmission electron microscopy (HRTEM) enables direct structural characterization, achieving atomic-scale resolution in HAP is challenging due to its beam sensitivity. Low-dose imaging mitigates beam-induced damage but often suffers insufficient contrast for local structural analysis. Herein, we developed an HRTEM imaging approach aided with single-image deep-learning denoising to investigate the structural effects of Na+ substitution in HAP. The denoising effectively removes noise from low-dose TEM images, facilitating both qualitative and quantitative analysis of atomic arrangements in HAP particles. We show that Na+ incorporation induces disordered surface layers, providing direct insight into ion-induced property modulation in HAP. Our low-dose imaging approach combined with single-image denoising offers a framework for atomic-scale structural characterization of beam-sensitive materials that are otherwise obscured by beam damage.

Image denoising remains a challenging problem due to the ill-posed nature of recovering clean images from noisy observations. Prior-based modeling plays a vital role in addressing this challenge. Low-rank priors effectively exploit non-local self-similarity by modeling correlations across similar images, but often over-smooth fine textures. In contrast, sparse priors, especially in the form of convolutional sparse coding (CSC), excel at preserving high-frequency texture details, yet typically neglect the structure across images. Recent unfolded CSC networks have improved denoising performance by combining CSC with deep networks, but most of them rely solely on sparsity, overlooking the complementary low-rank structure information. To overcome these limitations, we propose UCSC-LR, the first interpretable unfolded CSC network that jointly leverages both sparse and low-rank priors within a unified architecture. UCSC-LR unrolls both the Iterative Shrinkage-Thresholding Algorithm (ISTA) and Singular Value Thresholding (SVT) to strictly implement alternating minimization, enabling simultaneous pursuit of sparse representations and low-rank reconstructions. A dedicated Fusion Net is introduced to adaptively integrate features from the two priors, allowing the model to preserve both texture and structural content. To further enhance performance on color image denoising, UCSC-LR incorporates a lightweight channel attention mechanism to capture inter-channel dependencies. With shared, pre-learned convolutional dictionaries and efficient parameterization, UCSC-LR achieves state-of-the-art performance with low model complexity. Extensive experiments on grayscale, blind, and color denoising tasks validate the effectiveness of the proposed method, consistently surpassing existing CSC-based and deep learning-based denoisers.

Purpose.Diffusion-weighted imaging (DWI) has significant value in disease diagnosis and treatment response monitoring, but its inherent low signal-to-noise ratio (SNR) severely affects image quality and quantification accuracy. Existing denoising techniques often blur important tissue boundary information when suppressing noise.Methods.This study proposes a band-limited implicit neural representation (BL-INR) framework for DWI denoising. The method introduces BL positional encoding based on the frequency response characteristics of the sinc function to restrict INR models from learning high-frequency noise while maintaining strong signal representation capabilities. Furthermore, multi-b-value DWI and structural MRI from the same patient are integrated as anatomical priors, exploiting the correlation between true signals and the statistical independence of noise to achieve effective denoising.Main Results.In clinical DWI data evaluation across four anatomical regions (brain, head and neck, abdomen, and pelvis), BL-INR's visualization results were superior to existing methods. Under extremely low SNR conditions (SNR = 1) in simulated noise experiments, BL-INR achieved a peak SNR of 35.44 and structural similarity index of 0.933, significantly outperforming other methods. Phantom denoising results showed that BL-INR achieved an average apparent diffusion coefficient value error of only4.57×10-5 mm2 s-1, the smallest among all methods.Significance.BL-INR provides a novel approach for DWI denoising by limiting the frequency of INR input positional encoding. Its self-supervised learning characteristics require no paired training data and allow convenient clinical application. The method enables the derivation of accurate diffusion parameters, providing a reliable foundation for DWI-based quantitative analysis with significant clinical application value.

Anatomically and metabolically informed diffusion for unified denoising and segmentation in low-count PET imaging.

A denoising method based on total variation regularization was applied to X-ray diffraction images obtained from X-ray crystallography experiments. This approach significantly enhanced the signal-to-noise ratio of weak diffraction spots. Subsequently, the denoised images were used for crystal structure data processing using cytidine as a standard sample, which yielded an improved structural analysis result. This image processing technique offers a practical method for improving the quality of weak diffraction data, which is particularly relevant for challenging samples such as microcrystals, thereby enabling more reliable crystallographic analysis.

Denoising of radiation images: application on MOX fuel characterization

Amorphous materials-solids lacking long-range order-underpin technologies from thin-film electronics1, solar cells2 and phase-change memory3 to magnetic components4, medical devices5 and quantum technologies6-8. Yet the absence of periodicity fundamentally limits determination of their three-dimensional (3D) structure at atomic resolution. Despite major theoretical, experimental, and computational advances in characterizing short- and medium-range order9-24, quantitative determination of complete 3D atomic arrangements in amorphous materials remains experimentally demanding. Atomic electron tomography (AET) now provides a pathway to direct 3D atomic mapping in these materials25-27. Here we present a quantitative analysis of AET, showing how robust image preprocessing, denoising, projection alignment and normalization, advanced tomographic reconstruction, atom tracing, elemental classification and atomic position refinement collectively enable reliable determination of 3D atomic coordinates and elemental identities in amorphous materials. Using multislice-simulated datasets of amorphous Si, SiGeSn and CoPdPt nanoparticles under varying noise levels, our workflow outperforms an alternative approach28 in both positional precision and classification accuracy. For CoPdPt, we identify 95.1% of Co, 99.0% of Pd and 100% of Pt atoms, with corresponding 3D positional precisions of 29 pm, 12 pm and 6 pm, respectively, under realistic dose conditions. These results establish practical guidelines and quantitative benchmarks for achieving accurate AET of non-crystalline materials, and the underlying framework can be broadly applied to other tomographic imaging modalities for high-fidelity 3D reconstruction.

First-photon imaging denoising based on point cloud radius filtering

A multiplexed three-channel detection system for rapid home-based diagnosis of respiratory viruses.

Infrared image denoising: A hybrid noise removal pipeline with anti-artifact regularization

Cycle-constrained adversarial denoising convolutional network for PET image denoising: Multi-dimensional validation on large datasets with reader study and real low-dose data.

Multispectral filter array (MSFA) camera is increasingly used due to its compact size and fast capturing speed. However, because of its narrow-band property, it often suffers from the light-deficient problem, and images captured are easily overwhelmed by noise. As a type of commonly used denoising method, neural networks have shown their power to achieve satisfactory denoising results. However, their performance highly depends on high-quality noisy-clean image pairs. For the task of MSFA image denoising, there is currently neither a paired real dataset nor an accurate noise model capable of generating realistic noisy images. To this end, we present a physics-based noise model that is capable to match the real noise distribution and synthesize realistic noisy images. In our noise model, those different types of noise can be divided into SimpleDist component and ComplexDist component. The former contains all the types of noise that can be described using a simple probability distribution like Gaussian or Poisson distribution, and the latter contains the complicated color bias noise that cannot be modeled using a simple probability distribution. Besides, we design a noise-decoupled network consisting of a SimpleDist noise removal network (SNRNet) and a ComplexDist noise removal network (CNRNet) to sequentially remove each component. Moreover, according to the non-uniformity of color bias noise in our noise model, we introduce a learnable position embedding in CNRNet to indicate the position information. To verify the effectiveness of our physics-based noise model and noise-decoupled network, we collect a real MSFA denoising dataset with paired long-exposure clean images and short-exposure noisy images. Experiments are conducted to prove that the network trained using synthetic data generated by our noise model performs as well as trained using paired real data, and our noise-decoupled network outperforms other state-of-the-art denoising methods.

To address the prevalent noise issue in images generated by HumanNeRF, this paper proposes an image denoising method that combines self-supervised contrastive learning and Generative Adversarial Networks (GANs). While HumanNeRF excels in realistic 3D human reconstruction tasks, its generated images often suffer from noise and detail loss due to incomplete training data and sampling noise during the rendering process. To solve this problem, our method first utilizes a self-supervised contrastive learning strategy to construct positive and negative sample pairs, enabling the network to effectively distinguish between noise and human detail features without external labels. Secondly, it introduces a Generative Adversarial Network, where the adversarial training between the generator and discriminator further enhances the detail representation and overall realism of the images. Experimental results demonstrate that the proposed method can effectively remove noise from HumanNeRF images while significantly improving detail fidelity, ultimately yielding higher-quality human images and providing crucial support for subsequent 3D human reconstruction and realistic rendering.

Optical neuromorphic computing offers promising avenues for real-time image processing and low-power artificial intelligence. Here, we introduce and experimentally validate a fundamentally new computing paradigm that exploits optical exciton dynamics in two-dimensional van der Waals heterostructures. The type-II band alignment and high permittivity (ε ≈ 19) enable exciton control, achieving an exciton-to-trion ratio of ∼7 under electric field modulation at room temperature. Quasi-linear trion photoluminescence acts as an optical synaptic response, with weights dynamically tuned by substrate voltage. By leveraging these programmable optical responses, we have achieved neuromorphic functions, including convolutional filtering for image denoising and fully connected networks for pattern recognition, achieving a classification accuracy rate of 98.7% even under noisy conditions. This work establishes β-TeO2 as a key material for optical neural networks and adaptive vision systems, redefining intelligent photonic processing.

Abstract Noise is ubiquitous during image acquisition. Sufficient denoising is often an important first step for image processing. In recent decades, deep neural networks (DNNs) have been widely used for image denoising. Most DNN‐based image denoising methods require a large‐scale dataset or focus on supervised settings, in which single/pairs of clean images or a set of noisy images are required. This poses a significant burden on the image acquisition process. Moreover, denoisers trained on datasets of limited scale may incur over‐fitting. To mitigate these issues, a new self‐supervised framework for image denoising based on the Tucker low‐rank tensor approximation is introduced. With the proposed design, the authors are able to characterize the denoiser with fewer parameters and train it based on a single image, which considerably improves the model generalizability and reduces the cost of data acquisition. Extensive experiments on both synthetic and real‐world noisy images have been conducted. Empirical results show that the proposed method outperforms existing non‐learning‐based methods (e.g. low‐pass filter, non‐local mean), single‐image unsupervised denoisers (e.g. DIP, NN+BM3D) evaluated on both in‐sample and out‐sample datasets. The proposed method even achieves comparable performances with some supervised methods (e.g. DnCNN).

Low-dose computed tomography (LDCT) has been widely adopted in clinical imaging to reduce radiation exposure. However, the inherent quantum noise and streaking artifacts in LDCT markedly degrade image quality, thereby compromising diagnostic accuracy. While conventional model-based iterative denoising (MBIR) approaches effectively mitigate noise via rigorous physical modeling, their adoption is hindered by substantial computational overhead. Deep learning-based approaches demonstrate strong denoising capabilities but face challenges in generalizing across imaging scanners and protocols. In this study, we propose the Dual-Interactive Fusion Network framework (DIFNet) for LDCT images, integrating the Dual-Phase Denoising Architecture (DPDA), Context-Aware Training Strategy (CATS), and a combined dual-phase loss function. On two in‑house LDCT datasets acquired with different Philips scanners, our approach reduces noise and artifacts while preserving essential anatomical detail and outperforms established denoising methods such as REDCNN, EDCNN, DDPM, and CTformer in both qualitative evaluation and quantitative metrics. Evaluation on the public Mayo-2016 benchmark, collected using a Siemens scanner, confirms DIFNet's robust and competitive performance. Ablation experiments further validate the effectiveness of our overall network design and the contribution of its core components. Our findings highlight the potential of DIFNet for real-world clinical applications, improving diagnostic reliability and patient care. The proposed framework advances LDCT denoising by balancing performance, robustness, and computational efficiency.