Feng Yu

Objectives: This study aims to comprehensively assess the value of deep learning (DL) models in improving the image quality of respiratory-gated PET/CT and explore the impact of PET acquisition time on the quality of model-predicted images. Methods: We enrolled 90 patients who underwent respiratory-gated PET/CT whole-body scans, and the respiratory-gating trigger phase width was set to 35%. Using list-mode data from full-acquisition-time PET (FAT-PET, 240 s/bed), we reconstructed four sets of low-acquisition-time PET (LAT-PET) images at 30, 60, 90, and 120 s. We trained U-net and Generative Adversarial Network (GAN) models to enhance each dataset. The evaluation of denoising performance considered clinical aspects including subjective scores, metabolic metrics, lesion detection, and radiomic feature stability. Results: Both GAN and U-net models improved subjective Likert scores for image quality and lesion conspicuity while maintaining metabolic metric consistency. 120 s GAN and U-net enhanced images showed strong quantitative consistency with FAT-PET, with mean absolute percentage error (MAPE) of SUV _peak for lesions being 5.6% (bias, − 0.19 ± 0.20) and 5.5% (bias, − 0.09 ± 0.27), respectively. The detection of lesions in GAN- and U-net-enhanced images was positively correlated with acquisition time. GANs consistently enhanced lesion detection in LAT-PET images, whereas U-nets provided no benefit for 60 s or higher acquisition times. Compared with 120 s free-breathing PET (FB-PET), GAN-enhancement provided superior overall lesion detection at 90 s ( p = 0.035) and 120 s ( p < 0.001); 120 s LAT-PET also outperformed FB-PET ( p < 0.001), while 90 s LAT-PET and 120 s U-net-enhanced images showed comparable performance ( p = 0.749 and p = 0.086, respectively). We also found nine U-net enhanced images showed false-positive liver uptake, while GAN-enhanced images remained free of false positives. Additionally, most radiomic features from LAT-PET images showed greater consistency with FAT-PET features compared to those from AI-enhanced images. Conclusion: AI image enhancement—particularly generative models—can effectively reduce noise in LAT-PET images, improve lesion detection, and maintain metabolic metrics. However, they demonstrate limitations in restoring radiomic features. These findings validate the clinical benefits of AI-based PET image enhancement while acknowledging their constraints in radiomics analysis.

In this paper, we present a multi-parallel restricted additive Schwarz (MP-RAS) preconditioner construction method for cardiac electrophysiology simulation. This method is designed to address the need for solving large-scale linear systems in realistic cardiac electrophysiology simulations and can provide a more efficient computational tool for patient-specific electrical propagation modeling, arrhythmia studies, and the evaluation of ablation strategies. The proposed preconditioner is suitable for the finite element simulation of the anisotropic cardiac monodomain model. In particular, we construct the subdomains based on Morton code sorting, build submatrices by indices and decompose the formula for parallel computing. Given that the computing of each subdomain is relatively independent, the iteration can be extended to N-parallel. Numerical experiments indicate that for matrices of the same size and under the same number of partitions, Morton code sorting is at least 105 times faster than METIS, while the memory usages are reduced by 12∼32%. The iteration number is reduced by approximately two times compared with the Jacobi and block Jacobi preconditioned conjugate gradient (PCG) method. Comparative experiments with other solvers further demonstrate that the MP-RAS solver is highly efficient for solving this parabolic partial differential equation and have strong parallel scalability.

The hardware-based computation of arcsine and arccosine functions is predominantly realized through the vectoring-mode CORDIC algorithm. However, existing approaches necessitate computationally intensive calculations or suffer from approximation-induced errors. This brief proposes a hybrid architecture that replaces the sequentially dependent early iterations with an efficient table-based mapping. By reinterpreting the initial convergence process as a non-uniform domain segmentation problem, the proposed Prefix Binary Tree Mapping (PBTM) achieves algorithmic look-ahead through direct indexing of precomputed intermediate states, thereby bypassing the coarse-grained decision stages. Experimental results demonstrate that by substituting the early iterations with precise state mapping, the PBTM-CORDIC method suppresses approximate gain error below the quantization noise floor and reduces the worst-case error by 76.02%. Furthermore, hardware implementation confirms a 22.77% reduction in area and a 56.08% enhancement in energy efficiency compared to state-of-the-art pipelined designs.

Sleep micro-events, such as sleep spindles and K-complexes, are closely associated with neurological cognitive functions. While artificial intelligence (AI)-assisted sleep micro-event detection provides automated annotation to reduce reliance on labor-intensive expert labeling, current supervised approaches require precisely annotated datasets that remain scarce in clinical practice. To overcome this data bottleneck, this paper introduces a Weakly Supervised Sleep Micro-Event Detector (WSSMED) that leverages readily available coarse sleep stage annotations. The proposed WSSMED features a dual-branch architecture, consisting of a wave prototype module and a cluster module, designed to capture the fine-grained sleep micro-event patterns experts rely on for sleep staging. This framework infers expert logic from coarse annotations while mitigating performance degradation caused by annotation inconsistencies arising from inter-rater variability. Experiments conducted on two public datasets and one clinical dataset demonstrate that WSSMED achieves state-of-the-art performance in detecting sleep spindles and K-complexes, as evaluated at both sample-level and event-level in terms of precision, recall and F1-score metrics. Furthermore, subject-level evaluation demonstrates that the density and duration of micro-events detected by WSSMED-key metrics linked to cognitive function and neurological status-align more closely with expert annotations than those of other reported methods. These results highlight the clinical potential of WSSMED for reliable sleep micro-event analysis.

Introduction: Structural magnetic resonance imaging (sMRI) is an important tool for the early diagnosis of Alzheimer's disease (AD). Previous methods based on voxel, region of interests (ROIs) or patch have limitations in characterizing discriminative features in sMRI for AD as they can only focus on specific local or global features.Methods: We propose a computer-aided AD diagnosis method based on sMRI, named PlgFormer, which considers the extraction of both local and global features. By using a combination of convolution and self-attention, we can extract context features at both local and global levels. In the decision-making layer of the model, we design a feature fusion module that adaptively selects context features through a gating mechanism. Additionally, to account for changes in image input resolution during the downsampling operation, we embed a dynamic embedding block at each stage of the network, which can adaptively adjust the weights of the inputs with different resolutions.Results: We evaluated the performance of our method on dichotomous AD vs. normal control (NC) and mild cognitive impairment (MCI) vs. NC, as well as trichotomous AD vs. MCI vs. NC classification tasks, using publicly available ADNI and XWNI datasets that we collected. On the ADNI dataset, the proposed method achieves classification accuracies of 0.9431 for AD vs. NC, 0.8216 for MCI vs. CN, and 0.6228 for the AD vs. MCI vs. CN task. On the XWNI dataset, the corresponding accuracies are 0.9307, 0.8600, and 0.8672, respectively. The experimental results demonstrate the high precision and robustness of our method in diagnosing people with different stages of cognitive impairment.Conclusion: The findings in our experimental results underscore the clinical potential of our proposed PlgFormer as a reliable and interpretable framework for supporting early and accurate diagnosis of AD using sMRI.

In recent years, the continuous advancement of digital technologies such as telemedicine and medical cloud computing has promoted collaborative research and diagnosis across multiple medical centers. However, the timely remote transmission and analysis of large volumetric medical images still pose significant challenges. While classical methods predominantly employing lossless compression are increasingly constrained by the limits of compression ratios, lossy 3D medical image compression methods are emerging as a promising alternative. Different from the existing 3D convolutional compression algorithms oriented only for human vision, this paper proposes a Multi-scale Contextual Autoencoder (MCAE) architecture that recurrently incorporates anatomical inter-slice context to optimize the compression of the current slice for both human and machine vision. Our decoded intermediate features, with sufficiently preserved semantic information, enable high-quality visualization and allow downstream machine vision tasks (e.g., segmentation and classification) to be performed directly without pixel-level recovery. To reduce the compression bit cost, we create a Multi-Dimensional Entropy Model that integrates inter-slice latent context with spatial-channel context and hierarchical hypercontext. Experimental results demonstrate that our framework obtains an average 9% BD-Rate reduction over the Versatile Video Coding (VVC) anchor on MRNet datasets, while achieving superior recognition performance for downstream segmentation and classification tasks than inputting reconstructed lossy images.

Domain-specific computing architectures significantly enhance the scale and performance of scientific simulations. In molecular dynamics, optimization of interatomic interaction computations is crucial for simulation accuracy. However, current methods aimed at optimizing accuracy rely on predefined interaction pipelines whose static computational configurations often lead to compromised resource efficiency. To address the difficulty mentioned above in calculating interatomic forces, we propose a Distribution-Aware Pair-Interaction Calculation (DAPIC) architecture that dynamically senses atomic distribution and allocates computational resources adaptively. This DAPIC architecture assigns computational weights to critical pair-interactions, enhancing the accuracy of essential distance calculations. DAPIC integrates Taylor series-based interpolation with a dynamically reloadable lookup table, emphasizing high-impact interaction regions. The experimental results demonstrate that comparable computational accuracy is achieved with significantly reduced resource demands, alleviating performance bottlenecks in interatomic force calculations. Specifically, the DAPIC-based design reduced the cumulative interatomic force error per atom by 70.8% and mitigated the atomic position and velocity drift errors by 86.2% and 85.6%, respectively, compared with single-precision GROMACS simulations. In long-term energy conservation tests, the total energy fluctuations closely align with those observed in double-precision GROMACS, underscoring the effectiveness of the architecture in preserving simulation fidelity.