Individuals with Alzheimer's disease (AD) are known to have difficulty utilizing semantic strategy to remember items from a list, such as during word list recall tasks. However, it is unclear if presenting words in the same order as compared to shuffling across trials of a list-learning test confers an advantage in terms of word recall in individuals with AD. To address this issue, we leveraged data from 10,504 participants over five studies across four AD Research Centers that administered the Consortium to Establish a Registry for AD Word List Memory Test in either a standard sequence (shuffled words for each learning trial) or nonstandard sequence (unshuffled words in a consistent order for each learning trial). After controlling for gender, age, years of education, apolipoprotein E gene, Clinical Dementia Rating Global Score, and composite test scores for executive functioning and language comprehension, we found that differences in word recall performance between participants with and without probable AD were larger when they were tested with the unshuffled versus standard shuffled version of the word list. When administered the unshuffled Consortium to Establish a Registry for AD Word List, participants without AD demonstrated a larger recall advantage compared to those with probable AD. This advantage was higher on Trial 3 (β = 0.09) compared to Trial 2 (β = 0.04), suggesting that this advantage increases with multiple immediate repetitions. Using an unshuffled word list may enhance the ability to distinguish between individuals along the AD continuum. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
V(D)J recombination, which joins V, D, and J genes across immunoglobulin loci, is essential for antibody diversity. To unlock the inherent constraints of V(D)J recombination, we systematically designed, assembled, and shuffled a synthetic human immunoglobulin heavy chain locus (SynIgh). SynIgh was designed de novo, incorporating 60 VH, 31 D, and 13 JH from global populations, along with 63 synthetic rearrangement sites, endowing it with modularity, extensibility, and dynamicity. We then synthesized the repetitive SynIgh sequence with a stabilizer. Using this construct, we generated a SynIgh mouse model capable of undergoing V(D)J recombination. Moreover, we demonstrated that a dual-rearrangement strategy, combining synthetic rearrangement with natural V(D)J recombination, enables shuffling of the SynIgh locus in mice. This strategy expands the diversity of V(D)J recombination outcomes, thereby reconfiguring the process of antibody diversity generation. SynIgh establishes a generalizable framework for the design, building, testing, and learning of mammalian genome architecture and function.
Reading impairments are common in stroke-induced aphasia and limit participation in functional and leisure activities. Traditional rehabilitation strategies show limited generalization, underscoring the need for novel interventions targeting residual neural networks. This feasibility randomized controlled trial evaluated real-time functional magnetic resonance imaging (fMRI) neurofeedback intervention for poststroke reading deficits. Subacute left-hemisphere stroke survivors and healthy controls completed 3 weekly fMRI neurofeedback and 10 out-of-scanner practice sessions. Stroke participants were randomized to contingent neurofeedback (based on left supramarginal gyrus activity; n=4) or noncontingent neurofeedback (shuffled feedback from another participant; n=3). Healthy controls (n=4) received contingent neurofeedback and served as a normative reference. Primary outcomes were changes from baseline to postintervention (≈3 weeks) in task-based brain activity (motor imagery/word/nonword reading>baseline), resting-state connectivity, and reading aloud. Reading comprehension was a secondary outcome. Group×session effects were tested using repeated-measures analyses and planned contrasts. Task fMRI revealed training-related activation increases in the left supramarginal gyrus (z=4.7; cluster-corrected P=0.05) and broader reading network in the contingent neurofeedback group, particularly during nonword reading. Activation increases in the noncontingent stroke group and healthy controls were more widespread and less reading-specific. Resting-state fMRI revealed greater integration among motor, auditory, and language networks in the contingent groups, with more disorganized patterns in the noncontingent group (permutation P=0.01; Δr=-0.1 to 0.1). No changes were observed in reading aloud. A significant group×session interaction was found for Reading Comprehension Battery for Aphasia, second edition (F[2, 8]=8.00; P<0.05; η2=0.67). The contingent neurofeedback stroke group improved more than healthy controls (mean group difference in Reading Comprehension Battery for Aphasia, second edition, change score=9.75 [95% CI, 1.99-17.51]; t[6]=3.07; P<0.05) and more than the noncontingent neurofeedback stroke group (mean group difference in Reading Comprehension Battery for Aphasia, second edition, change score=11.42 [95% CI, 1.12-21.71]; t[5]=2.85; P<0.05). These findings support the feasibility of targeting the residual reading network during early recovery using fMRI neurofeedback. Confirmation of these preliminary effects awaits completion of the ongoing randomized controlled trial. URL: https://www.clinicaltrials.gov; Unique identifier: NCT04875936.
The automated inspection of aluminum profile surface defects, which heavily relies on data acquired by machine vision sensors, is a critical task in industrial quality control. Addressing the current challenges of intense background texture interference and the difficulty in detecting defects with extreme aspect ratios on aluminum profiles, this research puts forward a complete end-to-end defect detection algorithm named WMC-DFINE (WIFA-MKSS-CSFF-DFINE) based on the DFINE framework. First, a Wavelet-Integrated Frequency Attention (WIFA) module is introduced, which utilizes a discrete wavelet transform to decouple features into the frequency domain, thereby dynamically suppressing high-frequency background noise and enhancing defect edge responses. Second, a Cross-Scale Feature Fusion (CSFF) module based on dual-channel pooling is designed to ensure the continuity of defect features, thereby resolving the semantic misalignment issue in traditional fusion. Third, a Multi-Kernel Strip Shuffle (MKSS) module is incorporated, utilizing decomposed convolution kernels to capture the geometric features of slender scratches. Finally, a knowledge distillation strategy is employed to transfer structured knowledge from a complex teacher model to a lightweight student model. Experiments on the Tianchi aluminum defect dataset demonstrate that WMC-DFINE achieves a mAP of 82.1%, which surpasses algorithms including YOLOv12, RT-DETR, and the baseline model DFINE. Furthermore, the distilled student model, WMC-DFINE-distill, improves the mAP by 3.2% compared to DFINE, reduces parameter count by 47%, and achieves an inference speed of 59.75 FPS on the experimental equipment. The proposed method effectively resolves the problem of balancing background suppression and defect detail feature preservation, offering a practical and efficient scheme for real-time industrial defect inspection.
The effectiveness of multimodal face anti-spoofing largely depends on the modeling of cross-modal relationships. However, most existing approaches rely on static fusion or implicitly learned feature aggregation, which assumes fixed modality importance, limiting its ability to capture reliability variations across different attack patterns. Under strict computational constraints, achieving effective dynamic cross-modal modeling remains a significant challenge. To address this issue, we propose an ultra-lightweight dynamic cross-modal framework for face anti-spoofing, with ultra-low parameters, FLOPs, latency, memory and high FPS for real-time edge inference. A compact feature extractor is constructed by enhancing ShuffleNetV2 with the Ghost-Generated Shuffle BlockA (GGS-BlockA), which significantly reduces redundant computation while maintaining high discriminative capability. On this basis, a Lightweight Cross-Modal Attention (LCMA) module performs sample-wise dynamic modality reweighting to capture reliability variations among RGB, Depth, and IR modalities. Furthermore, a Lightweight Cross-Modal Fusion (LCMF) module utilizes depth cues as stable guidance to improve cross-modal feature alignment and complementary representation. Experiments on the CASIA-SURF benchmark demonstrate that the proposed method achieves an Average Classification Error Rate (ACER) of 0.064% with only 0.14M parameters and 0.0065G FLOPs. At the strict threshold of TPR@FPR=10-4, a detection rate of 99.86% is obtained, demonstrating strong robustness and generalization capability under extremely low computational cost.
Unlike two-dimensional van der Waals (2D vdW) materials, which achieve deformability through dimensional reduction, bulk materials rely more on their intrinsic crystal structures and chemical bond interactions for ductility. Considering the multi-scale complexity of plastic deformation and failure in bulk materials, we propose using B/G and κ criteria to synergistically pre-screen ductile materials. From 152 binary face-centered cubic (FCC) materials, we screened 16 compounds with good ductility, including zincblende structured FeN and rocksalt structured CdO. First principle calculations show that uniform deformation of the atomic framework is the main deformation mechanism for FeN and CdO. Both materials exhibit significant strain before failure, demonstrating good deformability. Shear-induced perfect dislocation slip of (111)/[Formula: see text] in the shuffle-set plane leads to the restoration and stress release of the FeN structure, thereby conferring potential ductility. For CdO, its two slip systems with lower ideal shear strengths are prone to simultaneous activation, and the < 001> direction exhibits a large tensile strain together with a low ideal tensile strength, both of which may promote its ductility. Generalized stacking fault energy (GSFE) calculations confirm the lowest energy barriers for the most active slip systems, consistent with the ideal shear strength results. This work provides a combined elastic criteria screening strategy and mechanistic insights into ductile binary FCC materials.
Understanding the physical and physiological demands of female youth basketball is essential for optimizing training and performance monitoring. However, evidence describing match demands in elite U-19 female players, particularly in African contexts, remains limited. Existing profiles are largely derived from male or adult cohorts and may not accurately reflect youth competition. This study aimed to examine the physical and physiological demands of elite North African U-19 female basketball players, considering differences by playing position (guards, forwards, centers) and competitive level (national vs. international). Thirty elite Tunisian U-19 female players (age 18.3 ± 0.2 years, height 1.78 ± 0.05 m, mass 82.9 ± 4.8 kg; 15 national-level, 15 international-level) were monitored during eight playoff games. Video-based time-motion analysis quantified activity frequency and duration across nine movement categories (standing, walking, jogging, running, sprinting, jumping, low/moderate/high-intensity shuffling). Physiological responses included heart rate (HR) monitoring (four intensity zones: <75%, 75-85%, 85-95%, >95% HRmax) and capillary [La] sampling. Two-way ANOVA (position × level) examined main and interaction effects. International-level players performed significantly more high-intensity activities than national-level players (224.0 ± 5.1 vs. 214.1 ± 5.4; p < 0.001, d = 1.95) and spent more time in maximal HR zones (16.1 ± 0.3% vs. 12.1 ± 0.3%; p < 0.001, d = 13.33), indicating greater fatigue resistance. Guards executed more high-intensity shuffling actions than forwards and centers (p < 0.001, η 2 = 0.92), whereas centers performed more static high-intensity actions and exhibited higher [La] concentrations (5.22 ± 0.15 vs. 4.93 ± 0.13 and 4.64 ± 0.12 mmol L-1; p < 0.001, η 2 = 0.76). High-intensity activity declined from the first to the fourth quarter in both groups (p < 0.001, η 2 = 0.94), with a greater reduction in national-level players (37.0% decline) than international players (31.9% decline). Intra-observer reliability was excellent across all movement categories (ICC ≥ 0.91; CV ≤ 4.6%). U-19 female basketball imposes distinct position-specific demands (guards: high-intensity lateral movements; centers: static exertions and elevated metabolic load) and competitive-level differences (international players: superior fatigue resistance). These findings provide the first quantitative profile of elite African female youth basketball, establishing reference benchmarks for position-specific conditioning and competitive-level progression assessment.
Dynamic facial expression recognition (DFER), which uses dynamic sequences (e.g., videos) of facial expressions rather than single static frames, incorporates temporal information about facial movements and thus greater ecological validity compared to static expression recognition. We studied DFER using a large, naturalistic database (Dynamic Facial Expressions in the Wild, DFEW) with over 11,000 video clips of 7 different facial expressions (Happy, Angry, Neutral, Surprise, Sad, Fear, Disgust). We created four modified versions of the database by varying Input type (Static cues or frames vs. Dynamic cues or optical flows) and temporal Structure (Ordered vs. Shuffled). This study examines and compares the facial recognition performance of deep convolutional neural networks (DCNNs) trained on these four Static and Dynamic datasets. Model performance was evaluated and compared across conditions using univariate and multivariate approaches. We found that DCNNs can recognize facial expressions from either Static or Dynamic cues. However, unlike in action recognition, Static cues consistently outperform Dynamic cues, perhaps due to the smaller magnitude of motion cues in facial expressions. Surprisingly, we found that disrupting temporal Structure did not affect the recognition of most facial expressions. Finally, multivariate analyses of the penultimate layers of the DCNNs suggested significant differences in the underlying representation of facial expressions between Static and Dynamic cues. Taken together, these findings provide new insights into the differential roles of Static and Dynamic cues in facial expression recognition.
Caspi et al. (see record 2026-80066-001) use the metaphor of "a deck of cards" in their work to describe psychopathology over the lifespan-specifically, how an individual's experience of psychopathology can change across development (both changing between different disorders and different presentations of the same diagnosis). Yet mental health research typically fragments these trajectories: Cross-sectional studies examine discrete time points, whereas longitudinal studies often track isolated disorders. Neither captures the dynamic, multidiagnostic life course patterns that Caspi et al. document. For example, someone may demonstrate anxious symptoms in adolescence, depressive symptoms in early adulthood, substance abuse in midlife, and psychosis symptoms even later in life. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
Group-aware learning has recently emerged as a promising paradigm for neuroimaging-based disease diagnosis, as population-level interactions can provide complementary information beyond individual imaging features. However, most existing approaches rely on explicitly constructed graphs, which introduce non-trivial design choices, scalability limitations, and sensitivity to graph topology. By incorporating the design philosophy of participatory interaction, we propose IP-Mamba, a scalable and memory-efficient framework tailored for neuroimaging cohorts that models implicit population interactions without the computational burden of explicit graph construction. IP-Mamba treats a mini-batch of subjects as an unordered set and employs a bidirectional Mamba-based sequence modeling mechanism to capture latent inter-subject dependencies. To address the inherent order sensitivity of sequence models, we introduce a Shuffle Consistency Strategy, which promotes permutation equivariance under random permutations of subject order, thereby aligning the model behavior with the clinically-relevant, set-based nature of population data. This design enables efficient implicit hypergraph modeling while maintaining linear computational complexity with respect to the population size. We evaluate IP-Mamba on the OASIS-1 dataset, focusing on the binary classification of Alzheimer's disease (Normal Controls vs. Abnormal) as an early clinical screening task. To address severe class imbalance and ensure diagnostic stability, we implement a Contextual Population Support Set inference mechanism coupled with a robust hybrid SVM decision layer. Experimental results demonstrate that IP-Mamba achieves a balanced accuracy of 87.84% and maintains a high sensitivity (Recall) of 89% for the minority disease class. Compared to conventional 3D CNNs and Transformer-based baselines, IP-Mamba provides highly competitive diagnostic robustness while maintaining a highly efficient linear O(N) memory scaling without the quadratic computational bottlenecks typical of graph-based attention networks. Comprehensive ablation studies further confirm the necessity of bidirectional modeling and shuffle consistency regularization. Overall, IP-Mamba offers a principled, memory-efficient alternative to explicit graph-based methods, providing a scalable solution for population-aware neuroimaging analysis under imbalanced clinical settings.
Soluble recombinant expression of Aedes aegypti mosquito midgut proteases in Escherichia coli prove to be difficult. These enzymes depend on disulfide bond formation for structural stability. Initial attempts in BL21(DE3) were unsuccessful due to a reducing cytoplasm. The use of T7 SHuffle cells (with a more oxidizing cytoplasm) led to soluble expression. However, other factors had to be altered (use of richer media and lower (< 25 °C) growth temperature). Not all mosquito proteases were equally soluble. Therefore, given the importance of IPTG in initiating transcription and translation, we set out to determine if low IPTG concentrations (< 0.1 mM) at 10 °C would increase soluble production of midgut proteases. Additionally, we investigated the effect of the small molecule osmolyte betaine on the soluble expression of midgut proteases. For this study, the focus was on Aedes aegypti Late Trypsin (AaLT), Early Trypsin (AaET), Serine Protease I (AaSPI), Serine protease V (AaSPV), and Juvenile Hormone Associated 15 (JHA15). The colder bacterial growth, along with low IPTG, slows the rate of transcription/translation of T7 RNA polymerase. Lower expression of T7 RNA polymerase, along with slower transcription activity at 10 °C, prevents rapid simultaneous translation of midgut proteases thereby promoting soluble expression. In addition, we found that different growth periods also varied among the proteases. Soluble expression for AaLT and AaET was maximal at 52 h post-induction, 72 h for JHA15, and 168 h for AaSPI and AaSPV. Surprisingly, for AaET, temperature was the only important factor. The addition of betaine to the growths had a more pronounced effect at higher (> 0.05 mM) IPTG. Low IPTG at 10 °C slows the rate of transcription/translation of recombinantly expressed mosquito proteases in bacteria. By preventing rapid accumulation in the cell, prevents aggregation, and ultimately inclusion body formation. Betaine works better at higher IPTG concentrations, but more studies are needed to better understand how this osmolyte stabilizes proteins during recombinant bacterial expression. Nonetheless, this study provides a blueprint for researchers who have never attempted IPTG concentrations < 0.05 mM to recombinantly express proteins in bacteria.
Deep learning has achieved remarkable progress in low-dose computed tomography (LDCT) denoising; however, radiologists struggle to trust black-box models they cannot verify or control. Zero-shot methods eliminate training data requirements but fail on computed tomography's (CT) spatially correlated noise. We demonstrate that a transparent mathematical operator, when made content-adaptive, can match deep learning performance while remaining fully interpretable. We introduce Filter2Noise (F2N), which replaces conventional deep networks with an attention-guided bilateral filter that adapts to local anatomy. A lightweight attention module (3.6k parameters) predicts optimal filtering strategies for each image region by analyzing tissue type, texture, and noise characteristics. To enable robust learning from a single noisy image with correlated noise, we develop Euclidean local shuffle, which strategically disrupts noise correlations while preserving anatomical structure, and a multi-scale self-supervised loss that enforces consistency across resolutions. On the Mayo Clinic LDCT Grand Challenge, F2N achieves 39.76 dB peak signal-to-noise ratio, outperforming the next-best zero-shot method by 1.88 dB, while using 360× fewer parameters (3.6k versus 1.3M). Clinical validation on photon-counting CT demonstrates that F2N elevates low-dose images to full-dose quality (no statistically significant difference in contrast-to-noise ratio, p = 0.10 ). The learned filtering strategy is fully visualizable: parameter maps reveal content-aware behavior. Radiologists can interactively adjust these parameters post-training to refine denoising in diagnostically critical regions. F2N reconciles competitive performance with complete interpretability and user control, providing radiologists with a verifiable tool that works across scanners and protocols without retraining.
Francisella tularensis is a Gram-negative bacterium which is highly infectious and poses a significant risk to public health as a Tier 1 Select Agent. Despite their essential roles in cell growth and division, the peptidoglycan cell wall and its associated remodeling enzymes remain poorly understood in F. tularensis. Lytic transglycosylases (LTs) are a class of enzymes critical for cell division and peptidoglycan recycling that represent attractive targets for novel antibiotics and inhibitors. Notably, the F. tularensis genome encodes only two LTs, MltA and Slt, both of which are essential for growth and contribute to virulence yet have not been biochemically characterized. In this study, we report the first expression, purification, and enzymatic characterization of recombinant His-tagged MltA (rMltA) and Slt (rSlt). rMltA was readily expressed in Escherichia coli and purified using affinity chromatography. However, rSlt proved refractory to conventional expression, and purification methods required extensive optimization to obtain soluble protein, including expression in SHuffle T7 E. coli and modified purification conditions. Both rMltA and rSlt were soluble in citric acid buffer at pH 5 and exhibited enzymatic activity in a fluorescent glycanase assay. rMltA displayed robust activity and quantitative kinetic analysis yielded a Km of 98.67 μM, whereas rSlt exhibited weaker but detectable activity. These results establish the first biochemical framework for studying the LT enzymes of F. tularensis and lay the foundation for future mechanistic studies and inhibitor screening to develop countermeasures for treatment of tularemia.
Nutritional deficiency in coffee is a major problem that compromises plant health, crop yield, and bean quality, directly threatening the economies of coffee-dependent regions. Traditional detection methods are primarily manual, time-consuming, and relied upon expert availability. This study introduces a novel Deep Learning (DL)-based dual-track architecture designed for the efficient classification of nutritional deficiencies in coffee leaf. The first track utilizes a MobileNetV3 backbone integrated with a Multi-Convolutional Shape-Aware Kernel (MCSK) block to capture spatially adaptive features from leaf textures and vein patterns. The second track employs a Hierarchical Shuffled Group Attention Network (HSGAN), utilizing Efficient Channel Attention (ECA) and Local Group Attention (LGA) modules to balance fine-grained local variations with broad spatial dependencies. Finally, a Multidimensional Collaborative Attention (MCA) mechanism is applied to the fused features to enhance cross-channel interactions and feature extraction. The proposed model was evaluated using the CoLeaf dataset, where it achieved an accuracy score of 96.04%. This performance demonstrates an improvement over existing research and current state-of-the-art models, highlighting the architecture's ability to identify complex nutrient-related patterns in coffee leaves. The performance of the proposed DL approach offer a solution for the automated monitoring of coffee plants. By providing a reliable alternative to manual inspection, this method presents the potential to help coffee production and support the agricultural regions worldwide.
Parkinson's disease (PD) presents significant challenges due to its intricate symptoms and often delayed diagnosis. Therefore, early detection is vital for effective management and slowing the disease progression. Recently, machine learning based methods show high performance in this purpose. This research introduces a new machine learning approach that combines Residual-Shuffle Network (ResNet) with an advanced metaheuristic, the Improved Dandelion Optimizer (IDO) by integrating adaptive parameter control and enhanced exploration-exploitation balance, to provide an accessible and precise solution for automated PD detection. The proposed framework addresses previous limitations by effectively adjusting model hyperparameters and network weights without the need for expensive or sophisticated data collection devices. The proposed IDO-ResShuffle framework achieved strong performance on the HandPD dataset, obtaining 97.6% accuracy, 96.9% F1-score, 97.2% sensitivity, and 97.1% specificity. These results demonstrate the effectiveness of jointly optimizing the network architecture and hyperparameters through the Improved Dandelion Optimizer, enabling more reliable identification of Parkinson's disease from handwriting patterns. These enhancements empower healthcare professionals to make informed decisions about patient care sooner and potentially slow down the progression of symptoms. By enhancing accessibility and reliability, this approach can enhance clinical decision-making and support timely intervention in PD management.
Protein N-glycosylation is a crucial post-translational modification that regulates cellular processes such as cell signalling, development, and autophagy. Abnormalities in protein glycosylation can manifest in life-threatening conditions, such as cancer. N-glycan analysis is crucial for determining the underlying cause of disease. The characterization of N-linked glycans is achieved through the removal of the carbohydrate moiety using deglycosylating enzymes. Peptide-N-glycosidase F (PNGase F) is a glycoamidase that hydrolyzes the amide bond in glycosamide by specifically cleaving at the innermost N-acetylglucosamine (GlcNAc). Although PNGase F has significant therapeutic applications, its widespread commercial use is limited by its high cost. This study focused on heterologous expression of Flavobacterium meningosepticum PNGase F in E. coli BL21 (DE3) and SHuffle® cells, in which the protein was partially soluble. The E. coli SHuffle® cells yielded 210.41 mg/L of PNGase F in TB glycerol medium in shake flasks, with a corresponding YP/X of 47.20 mg/g DCW. Expression studies in E. coli BL21 (DE3) cells yielded inclusion bodies (IBs) that were solubilized, yielding activity comparable to that of the soluble form. Furthermore, the fed-batch cultivation in E. coli BL21 (DE3) cells produced 5.87 g/L of IBs with a final OD600 of 176. Therefore, this study investigates the potential of alternative E. coli hosts as cost-effective production platforms for PNGase F.
RNA polymerase II (RNAPII) is a conserved 12-subunit enzyme essential for eukaryotic transcription. Although the structure and biological functions of RNAPII are well-defined, the mechanisms by which its subunits are assembled into a functional complex remain only partially understood. Several RNAPII assembly factors have been identified, but the molecular principles by which they cooperate during polymerase biogenesis remain unclear. Rba50 is an essential, conserved RNAPII assembly factor implicated in multiple stages of polymerase biogenesis, yet its mode of action remains unknown. Here, we combine structural prediction, targeted interaction assays, and functional analyses to reveal a modular organization of Rba50 with separable activities. Rba50 comprises two functionally distinct but cooperative regions: a structured C-terminal module that provides RNAPII assembly-associated activity and an intrinsically disordered N-terminal module that functions as a multivalent interaction hub. We map a direct interaction between the assembly factor Npa3 and the first N-terminal disordered segment, residues 1-89, and show that the N-terminal module expands the Rba50 interaction network. Functional assays indicate that the C-terminal region rescues specific rba50-3 phenotypes and supports proliferative recovery under permissive plasmid-shuffle conditions, whereas robust Rba50 function also requires determinants outside the C-terminal region. We further identify a CRM1-dependent leucine-rich nuclear export signal within residues 239-249 that limits nuclear accumulation of Rba50 at steady state. Together, our results support a modular architecture in which distinct functional elements within Rba50 contribute to RNAPII biogenesis, interaction-network formation, and nucleocytoplasmic distribution.
Deep neural networks are vulnerable to transferable adversarial examples in black-box scenarios, and targeted attacks that mislead models into predicting specific classes pose particularly severe threats. Feature mixup attacks enhance adversarial transferability by injecting clean features to perturb intermediate representations, yet existing methods suffer from two fundamental limitations: coarse-grained global mixing strategies apply a shared mixing ratio uniformly across all spatial positions, fixing the clean reference for each position to its corresponding location in the clean image and thus limiting the diversity of mixed feature representations during optimization and the transferability of the generated adversarial examples; and standard momentum-based optimization over-aligns with the surrogate model's gradient geometry, suppressing gradient variations essential for escaping model-specific local minima. We propose Fine-grained Feature Mixup Perturbation and Reference-based Gradient Refinement (FMGR) to address both limitations. FFM partitions feature maps into spatially disjoint blocks and mixes each block with clean features drawn from spatially shuffled positions of the same image, breaking the fixed spatial correspondence of clean references and producing more generalizable feature perturbations. RGR selectively amplifies deviations between instantaneous and reference gradients, suppressing surrogate-specific dominant directions and steering optimization toward flatter, more transferable loss regions. Extensive experiments on ImageNet and CIFAR-10 demonstrate that FMGR significantly outperforms state-of-the-art methods against both CNN and Vision Transformer architectures while maintaining computational efficiency.
Printed circuit board (PCB) defect detection faces significant challenges including dense micro-scale targets, similar background textures, and parameter redundancy in existing models. This study proposes MicroDETR, a lightweight real-time detection transformer optimized through three innovations. First, an EfficientBlock backbone employs depthwise separable convolution and channel shuffle mechanisms to enhance feature representation while compressing parameters. Second, an Enhanced Multi-Scale Feature Fusion Network (EMSFN) integrates multi-level features through bidirectional information flow. Third, a Multi-Level Feature Aggregation Module (MLFAM) reconstructs the AIFI component using grouped parallel computation to reduce complexity and enhance micro-target responses. Experimental results demonstrate that MicroDETR achieves 5.1% and 2.5% improvements in mAP@0.5 and mAP@0.5-0.95, respectively, with 3.3% and 0.8% enhancements in recall and precision. Simultaneously, the model reduces parameters by 15.21%, decreases computational complexity by 12.11%, and improves processing speed by 0.8 milliseconds. Comprehensive evaluation across seven defect categories reveals exceptional performance on small-scale targets (65% of dataset), with mAP@0.5 exceeding 77.5% for densely distributed defects. Robustness analysis demonstrates superior noise resistance, maintaining 78.1% mAP@0.5 under boundary noise (2.2% degradation vs. 5.6% for baseline) and achieving 69.8% mAP@0.5 with only 10% training data. Successful deployment on NVIDIA Jetson TX2 validates real-world applicability, delivering 18.9 FPS with 79.8% mAP@0.5 through TensorRT-FP16 optimization. Validation on chip surface defect and PKU-Market-PCB datasets confirms the effectiveness of the proposed approach in achieving optimal balance between accuracy and efficiency for industrial defect detection applications. The code and trained models are publicly available at https://github.com/yixing166/MicroDETR.
Real-world diabetic retinopathy (DR) screening faces a paradox: the most diagnostically critical images are often the lowest in quality, because advanced disease itself produces vitreous hemorrhage, proliferative tissue, and media opacities that degrade fundus imagery. We characterize this quality-severity coupling quantitatively (Spearman ρ = 0.420, odds ratio 4.17 for referable DR in Reject vs. Good strata on DDR, p < 0.001) and show that conventional pipelines work against it: filtering low-quality images discards the most severe cases, while uniform processing leads to misclassification. Both behaviors stem from treating image quality assessment (IQA) as a binary preprocessing decision. We argue that quality should serve as a continuous guidance signal that conditions the diagnostic process, and propose QGDR, a quality-guided dynamic routing framework realizing this paradigm through three coordinated mechanisms: (i) a multi-level IQA module that extracts hierarchical quality features across backbone stages; (ii) a quality-conditioned context gating mechanism that modulates spatial attention according to the predicted quality state; and (iii) an adaptive gated fusion mechanism that routes inputs to scale-specialized experts, with high-quality images preferentially activating fine-scale experts for subtle lesions and degraded images relying on coarse-scale experts for robust global pattern recognition. On EyeQ and DDR, QGDR attains 78.32% accuracy with 0.6863 QWK and 80.85% accuracy with 0.8231 QWK respectively, outperforming representative CNN, transformer, and foundation-model baselines while remaining within the compute envelope of standard single-stream backbones. Counter-intuitively, performance is preserved or improved on the lowest-quality stratum (77.61% on EyeQ-Reject; 82.21% on DDR-Reject, exceeding the Good-quality accuracy on the same dataset), and a Shuffled-IQA counterfactual confirms that QGDR exploits the semantic content of quality information rather than a generic auxiliary signal. Test-only evaluation on the external IDRiD and DeepDRiD cohorts confirms cross-dataset generalization. By treating image quality as guidance rather than as a filter, QGDR preserves screening coverage without sacrificing diagnostic reliability.