TL;DR
A novel cross-domain DL framework for MRI reconstruction that leverages graph-based convolutional neural networks to model the autoregressive nature of Fourier features and integrates spatial domain networks and FDNs through a cross-lattice structure, enhancing feature extraction by promoting dense representation through alternating layers.
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.
Results indicate that DMFT provides a promising framework for accurate and efficient cardiac MRIreconstruction, and structure aware prior guidance offers additional interpretability support in Healthcare 4.0 scenarios.
Deep learning–based multi-contrast Magnetic Resonance (MCMR) super-resolution (SR) has achieved notable success in accelerating image acquisition and improving image quality. However, significant challenges remain when dealing with the volumetric data: 1) Most existing MCMR SR methods primarily rely on single-slice information and fail to exploit high-dimensional volumetric contextual information; 2) Due to the sparsity of the original low-resolution volumetric data, conventional small kernel convolutions struggle to capture long-range contextual information. Although transformer-based approaches can model long-range dependencies, they suffer from high computational and memory demands when applied to high-dimensional volumetric data. To address these challenges, we propose a multi-view large-kernel attention network for MCMR volumetric SR. The method contains three stages: a cross-modality synthesis stage, an inter-slice deformable compensation stage, and a multi-view large-kernel attention fusion stage. Specifically, a multi-view fusion strategy is proposed to exploit the rich spatial contextual information inherent in high-dimensional volumetric data. A large-kernel convolution attention block is proposed to efficiently capture long-range dependencies from the sparsely sampled coronal and sagittal planes. By jointly integrating the high-order multi-view and multi-contrast information, our method successfully reconstructs high-quality MCMR volumetric data. Experimental results across different datasets, along with the downstream segmentation tasks, attest to the effectiveness of the proposed method.
Background and Objective Positron Emission Tomography (PET) is widely used for assessing metabolic activity and diagnosing cancer. Due to the inherently high noise levels in PET data, advanced reconstruction algorithms are essential for accurate imaging. Convolutional Neural Networks (CNNs) within the Learned Primal-Dual (LPD) framework have shown good performance for PET reconstruction; however, the locality assumption imposed by CNNs can limit their ability to capture long-range contextual dependencies in sinogram data. Motivated by our previous work on transformer-based architecture for low-dose and low-count sinogram denoising, we comprehensively investigate the integration of attention mechanisms, both self-attention and cross-attention, within the LPD framework. Methods We propose and evaluate three novel transformer-based LPD architectures: the Dual-Domain Stacked Transformer-based LPD, which employs sinusoidal patches to model long-range dependencies in the dual domain (sinogram); the Dual-Domain Restormer-based LPD, a hybrid design that combines CNNs for fine-grained local feature extraction with transformers for global information exchange; and the Dual-Domain UNet-based LPD incorporating Cross-Attention mechanism, which augments the U-Net LPD with cross-attention to capture complementary information across image and sinogram domains. To enable robust transfer learning, we further introduce a system-aligned synthetic data generation process that replicates the MiniPET-3 system, our in-house preclinical PET scanner dedicated to small-animal imaging, incorporating realistic acquisition geometry, resolution blurring, positron range effects, and Poisson noise. Results Evaluations on synthetic data demonstrate that both the Stacked Transformer-based LPD and Dual-Domain Restormer-based LPD outperform the UNet-based LPD baseline in MSE and PSNR across all noise levels, with the Restormer additionally achieving higher SSIM at low noise. The UNet-based LPD incorporating a cross-attention mechanism also improves MSE and PSNR over the baseline, while exhibiting a slight reduction in SSIM, indicating a trade-off between pixel-wise accuracy and structural similarity. Trained exclusively on synthetic data, the proposed models generalize well to experimental measurements, producing high-quality reconstructions. Conclusions Overall, this work highlights the promise of transformer-integrated LPD frameworks for enhancing PET reconstruction quality and robustness in low-count imaging settings.
COJEPA is presented, a self-supervised framework for volumetric brain MRI that combines a joint-embedding predictive architecture (JEPA) with a contrastive loss (CO), targeting two complementary properties: local predictivity and global discriminability.
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.
