2026· International Journal of Advanced Computer Science and Applications· 0 citations· 46 references
TL;DR
The iterative three-operator splitting (TOS) method to solve the composite regularization problem and a composite denoising network (CDNet) that maintains consistent arithmetic structures with it are proposed, demonstrating the effectiveness of deep unrolling in accelerating iterative optimization.
Abstract
Magnetic resonance imaging (MRI) is an essential modality in contemporary medical diagnosis, offering high-resolution images with excellent soft-tissue contrast of internal organs in a non-invasive, non-ionizing, and non-carcinogenic manner. However, the primary drawback of this technique is a lengthy data-acquisition process, which raises the possibility of picking up artefacts as well as discomfort in patients. To accelerate it, data is acquired at sub-Nyquist rates, and the image is recovered from undersampled, noisy measurements using the compressed sensing (CS) technique. Composite priors in CS–MRI have demonstrated superior performance compared to single priors, despite having limited image priors, which also leads to slower reconstruction speed. To date, deep unfolded networks (DUNs), which integrate the powerful data-driven prior of traditional deep learning (DL) with the strong interpretability of optimization algorithms, are the most outstanding reconstruction technology in MRI acceleration, yielding superior and interpretable results. This study proposes the iterative three-operator splitting (TOS) method to solve the composite regularization problem and a composite denoising network (CDNet) that maintains consistent arithmetic structures with it. The CDNet is implemented using complex-valued DL strategies for richer representation of MRI. Extensive experiments using brain and knee raw k-space from the FastMRI dataset with acceleration factors (AF) ranging from ×2 to ×10 and assessed using peak signal–to–noise ratio (PSNR), structural similarity index (SSIM), and normalized root–mean–squared error (NRMSE) validate the CDNet. Across AFs, CDNet achieved dataset-averaged PSNR/SSIM/NRMSE of 26.50–33.47dB/0.7787–0.8992/0.2562–0.1303 for brain MRI and 28.08–35.52dB/0.7069–0.8734/0.1929–0.0950 for knee MRI. The proposed network yields superior reconstruction accuracy and faster reconstruction speed compared to the iterative TOS for composite regularization, demonstrating the effectiveness of deep unrolling in accelerating iterative optimization. Furthermore, CDNet outperforms other state-of-the-art DUNs in reconstruction accuracy at comparable inference speed, demonstrating its vast potential for improving patient care, scanner throughput, and healthcare economics.
Low-dose computed tomography (LDCT) is a significant non-invasive imaging modality for disease diagnosis in early stages and clinical oncology. However, the reduction of radiation dose unavoidably introduces severe quantum noise, photon starvation and Poisson-Gaussian noise, which degrade contrast-to-noise ratio (CNR) and obscure subtle anatomical details. Recent advances in Artificial Intelligence (AI) have shown great promise in medical image restoration. However, pure deep learning methods often suffer from over-smoothing of fine structures and poor interpretability, while traditional non-convex variational models can preserve global edges, but are sensitive to the choice of parameters and produce staircasing artifacts. We propose an AI-empowered hybrid restoration framework that combines non-convex Total Variation (TV) optimization and a deep convolutional residual network within the Plug-and-Play (PnP) Alternating Direction Method of Multipliers (ADMM) framework to enjoy the complementary merits of the two paradigms. The AI based deep residual network can learn complex noise features and image priors efficiently. The optimization part keeps the structural fidelity and ensures the stable reconstruction. The proposed framework is tested on clinical lung CT slices from LIDC-IDRI benchmark dataset, and the experimental results show that the proposed framework achieves 33.97dB of Peak Signal-to-Noise Ratio (PSNR) and 0.918 of Structural Similarity Index Measure (SSIM) at noise level of σ=25. The experimental results show that the proposed AI-enabled hybrid model can better preserve structure edges, recover fine anatomical details and suppress noise compared with the traditional optimization methods and deep learning alone, which shows the effectiveness for low-dose medical image denoising.
M. Kristappa, Krishnanaik Vankdoth· International journal of com...· 0 citations
Self-supervised methods showed superior performance over supervised and classical methods for low-field knee MR images, and validation in a larger dataset of knee at 0.55T is needed to further support this conclusion.
Biomedical imaging systems typically operate under low signal conditions. As a consequence, the acquired measurements are often affected by Poisson noise, which arises from the stochastic nature of photon counting and varies with the underlying signal intensity. Although supervised deep learning approaches have achieved strong performance in image denoising tasks, obtaining clean reference images (ground truth) is rarely feasible in real clinical settings. In this work, we consider the Poisson2Sparse framework, which combines sparse representation theory with a deep algorithm unrolling architecture and enables self-supervised learning directly from noisy observations. While most existing implementations rely on CUDA-based GPU environments, the present study re-implements and evaluates the framework on Apple M4 hardware using the Metal Performance Shaders (MPS) backend. Experiments conducted on the PINCAT dataset show that the proposed implementation attains a PSNR of 31.88dB at a photon intensity level of 20, corresponding to a medium-noise regime and closely aligning with previously reported baseline results. These findings suggest that advanced self-supervised restoration models can be executed reliably on consumer-grade hardware equipped with a unified memory architecture (UMA), making such approaches more accessible for both research and potential clinical deployment.
Abdullah Altepe, Rıfat Volkan Şenyuva· Signal Processing and Commun...· 0 citations
A model-driven bilevel optimization framework that couples SENSE-based image reconstruction with SPIRiT-based k-space calibration through shared CSMs, and introduces a deep-prior-guided regularization strategy that preserves the structure of classical linear regularizers while adaptively learning spatially varying regularization weights from denoised intermediate reconstructions.
Weipeng Chen, Yan-Ran Li, Raymond H. Chan et al.· Journal of Mathematical Imag...· 0 citations
This work designs an implicit pixel-wise learnable step size to adapt to the spatial gradient heterogeneity of CT images and develops a cross-prompt guiding mechanism to enable inter-domain prompt interaction, which facilitates efficient prompt generation and enhances the convergence stability of the model.
Wenchao Du, Qiao Mu, Huanhuan Cui et al.· IEEE Transactions on Medical...· 0 citations
Low-field magnetic resonance imaging (MRI) is an affordable medical imaging technique used to assess the structural and functional features of internal and external tissues and organs. However, its clinical effectiveness is often constrained by prolonged scan times and reduced image quality due to compromised signal-to-noise ratios. Deep learning (DL) has emerged as an effective solution for reconstruction of undersampled MRI data. This study presents a comprehensive comparative evaluation of six U-Net variants, including U-Net, Attention U-Net, Residual U-Net, Recurrent-Residual (R2) U-Net, Residual-Attention U-Net (RAU-Net), and U-Net++, for image domain reconstruction of low-field (0.3 Tesla) MRI data from the M4RAW dataset. Root Sum of Squares (RSS) magnitude images were used for training and testing. The models were trained on 1024 MRI volumes comprising 18,432 image slices. Reconstruction performance was evaluated under Cartesian undersampling at acceleration factors of 2, 4, 8, and 16, and estimated different statistical indices including Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Normalized Mean Squared Error (NMSE) and Coefficient of Determination (R2). The results demonstrate that the vanilla U-Net performs well at lower acceleration factors, achieving average SSIM and PSNR values of 0.9303 and 39.18 dB, respectively, at an acceleration factor of 2. In contrast, more complex architectures, particularly U-Net++ and RAU-Net, exhibit superior performance at higher acceleration factors. In particular, U-Net++ achieved average SSIM values of 0.8754 and 0.8731, and PSNR values of 35.21 dB and 35.06 dB at acceleration factors of 8 and 16, respectively. These findings inform the development of efficient and accessible imaging solutions, promote the broader adoption of low-field MRI, and encourage further innovation in cost-effective medical imaging technologies.
Akif Ahmed Nasif Purno, K. M. T. K. Siddiki, S. M. Chapal Hossain· Magnetic Resonance Imaging· 0 citations
