The performance and stability of pharmaceutical hydrogels depend on microstructural features such as particle size and agglomeration. Characterization of these features is commonly performed through manual microscopic assessment, which can be subjective when particle boundaries are indistinct or agglomerates lack clearly defined interfaces. This study applies computer vision to automate the analysis of hydrogel micrographs, including particle segmentation, agglomerate identification, and the extraction of size-related morphological parameters. Two deep instance segmentation networks, the Mask R-CNN and the Mask2Former, were applied for this purpose. This study introduced a particle dilation and a graph-centric approach for agglomerate identification, differing from prior methods where agglomerates were annotated individually, thus mitigating some annotation challenges. The networks, trained on a hydrogel database, exhibited an Average Precision of 92.47% for Mask R-CNN and 91.86% for Mask2Former. The Mask2Former demonstrated superior Average Recall (AR) at 76.6%, compared to Mask R-CNN's 72.32%. This study pioneers the application of Mask2Former, a Vision Transformer-based network, for particle segmentation which had superior AR performance. Considering the subjective nature of annotation for hydrogel micrographs, where false positives can be considered as valid particles, this study recommends the inclusion of AR as a metric for model selection. Furthermore, the extracted morphological features from segmented images showed close agreement with manual measurements. This workflow has potential to support formulation development and quality assessment in pharmaceutical settings.

Efficient FPGA realization of neural network activation functions using adaptive piecewise polynomial approximations with Chebyshev nodes

Adversarially robust knowledge distillation aims to compress a large-scale robust teacher model into a lightweight student counterpart while preserving adversarial robustness and natural performance. Previous methods primarily focused on aligning knowledge (e.g., predictions) between teacher and student models to transfer robustness. However, potentially incorrect predictions from the teacher can misguide the student, negatively impacting robustness transfer. To circumvent this, we propose a novel adversarially robust knowledge distillation scheme that promotes alignment towards more benign predictions rather than incorrect ones by refining inputs into so-called "inverse adversarial examples" via simply reversing the sign of adversarial perturbation. Through a comprehensive investigation of the properties of inverse adversaries, we provide new theoretical insights showing how mimicking the behavior of the teacher model on inverse adversaries facilitates reliable robustness transfer built upon the implicit connection between robustness and the input gradient information. We thus design a gradient matching mechanism between teacher and student models utilizing inverse adversaries to facilitate robust knowledge alignment. Furthermore, inspired by our analysis of the correlation between robustness and adversarial transferability, we propose a weight-space disruption strategy that jointly interacts with both teacher and student models to find a shared direction for better robustness transfer. Empirical evaluations across various datasets demonstrate that our method achieves state-of-the-art robustness and natural performance. Notably, on ImageNet, our approach outperforms prior methods by approximately 3.8% in both clean and robust accuracy. Moreover, we show that incorporating auxiliary generated data into distillation further boosts robustness. Our method can also be generalized to multimodal architectures.

Accurate assessment of surgical complexity is essential in laparoscopic cholecystectomy (LC), where severe inflammation is associated with longer operative times and increased risk of postoperative complications. The Parkland Grading Scale (PGS) provides a clinically validated framework for stratifying inflammation severity; however, its automation in surgical videos remains largely unexplored, particularly in realistic scenarios where complete videos must be analyzed without prior manual curation. In this work, we introduce STC-Net, a novel framework for Single-Timestamp-based Complexity estimation in LC via the PGS, designed to operate under weak temporal supervision. Unlike prior methods limited to static images or manually trimmed clips, STC-Net operates directly on full videos. It jointly performs temporal localization and grading through a localization, window proposal, and grading module. We introduce a novel loss formulation combining hard and soft localization objectives and background-aware grading supervision. Evaluated on a private dataset of 1859 LC videos, STC-Net achieves an accuracy of 60.18% and an F1-score of 59.68%, outperforming non-localized baselines by over 12% in both metrics and highlighting the effectiveness of weak supervision for surgical complexity assessment. STC-Net demonstrates a scalable and effective approach for automated PGS-based surgical complexity estimation from full LC videos, making it promising for postoperative analysis and surgical training.

A sealed-vessel approach produced high-purity wurtzite hexagonal ZnO nanorods (P63mc, JCPDS 00-36-1451) with an exceptionally rapid holding duration of 50 s at 150 °C, utilizing microwave radiation from a Monowave 400 reactor. The process is more rapid than prior microwave methods, which require 2-7 min, and is significantly quicker than hydrothermal and sol-gel techniques that take hours to days, utilizing less than 10 Wh of energy compared to the energy-intensive hydrothermal method (energy savings exceeding 95%). The XRD study indicated a Scherrer crystallite size of 25.22 nm, a crystallinity of 48.5%, and a microstrain, as per the Williamson-Hall equation, of ε × 103 = 3.39. The results are consistent with a highly organized hexagonal wurtzite structure characterized by cell parameters a = b = 3.2494 Å and c = 5.2066 Å. A TEM study of 100 nanorods revealed a uniform morphology with an average diameter of 22.4 ± 3.2 nm, a length of 185 ± 25 nm, and an aspect ratio of 8.3 ± 1.4, indicating preferential development along the c-axis attributable to microwave coupling. The UV-Vis spectrophotometer yielded a cutoff absorption wavelength of λ = 374 nm, which corresponds to a band gap energy of 3.318 eV. The Fourier transformed infrared (FTIR) spectra validated the lattice vibrations of Zn-O bonds at 436 and 630 cm-1. Energy dispersive X-ray (EDX) revealed a near-stoichiometric composition (Zn = 51.48% and O = 48.52%), with no detected contaminants.

Denoising images captured under extreme low-light conditions remains a persistent challenge in computational photography, primarily due to low signal-to-noise ratios and sensor-specific noise characteristics. These variations often require persensor noise calibration to achieve effective denoising. Although recent calibration-free methods aim to reduce this dependency through synthetic noise modeling or few-shot fine-tuning, their performance often degrades in extreme low-light scenarios across different sensors due to mismatches between synthetic and real-world noise. To address this gap, we introduce CLIP-Guided Denoising (CLD), the first framework to leverage large-scale vision models pretrained on sRGB images for cross-domain feature fusion, effectively guiding RAW image denoising across diverse sensors. Although not trained on RAW data, CLIP embeddings offer semantically robust and noise-invariant features that help guide the denoising network to focus on the underlying image content rather than fitting to specific noise distributions. Extensive experiments on the SID and ELD datasets demonstrate that CLD achieves state-of-the-art performance in calibration-free settings, significantly outperforming prior methods under extreme low-light conditions and achieving robust generalization across unseen sensor domains.

Accurate and reliable estimation of the State of Health (SOH) of lithium-ion batteries is a fundamental requirement for the safe operation of Battery Management Systems in electric vehicles. Existing methods are broadly limited by the underutilization of complementary multi-modal information, overly simplistic fusion strategies that neglect cross-modal interactions, and the absence of explicit modeling between macroscopic temporal degradation signals and microscopic electrochemical impedance parameters. To address these limitations, this paper proposes DS-Transformer (Dual-Stream Transformer), an end-to-end dual-stream Transformer network. The model processes heterogeneous inputs through two parallel encoding paths: Stream1 employs a multi-scale one-dimensional convolutional neural network to extract local degradation features from voltage, current, and temperature time-series signals recorded during discharge; Stream2 uses a multi-layer perceptron to semantically encode the internal resistance ([Formula: see text]) and charge transfer resistance ([Formula: see text]). Unlike prior multi-modal methods that rely on simple feature concatenation, the two streams are dynamically aligned and explicitly fused through the proposed Cross-Stream Cross-Attention (CSCA) module, which uses temporal embeddings as Query and impedance embeddings as Key/Value to achieve semantic-level interaction modeling between degradation patterns and electrochemical states. A Transformer Encoder then captures global sequence dependencies, ultimately yielding SOH regression predictions. Systematic validation is conducted on the publicly available NASA AMES lithium-ion battery aging dataset (34 batteries, 7,565 experiments, covering five temperature conditions from 4[Formula: see text]C to 44[Formula: see text]C). DS-Transformer achieves MAE=1.24%, RMSE=1.67%, and [Formula: see text]=0.9782, improving over the best baseline by 28.3%, 27.1%, and 1.90%, respectively. Ablation studies quantitatively demonstrate that the CSCA module is the most critical contributor (MAE increases by 37.9% upon removal, outperforming simple concatenation by 27.5%), while the dual-stream architecture and Transformer encoder are each independently indispensable. Multi-temperature robustness experiments further validate the stable generalization capability of the proposed model. This work establishes, for the first time, an explicit cross-modal interaction framework between discharge temporal signals and electrochemical impedance parameters for battery SOH estimation, offering significant theoretical insights and practical engineering value.

Dementia, particularly Alzheimer's disease (AD), is a growing concern in aging populations, with mild cognitive impairment (MCI) frequently progressing to AD. While existing studies often rely on comprehensive neuropsychological evaluations and assessments by neurologists and neuroradiologists, these approaches are not always feasible in routine or rural clinical practice. Current diagnostic methods rely on clinical assessments and MRI-based visual scoring systems such as MTA, ERICA, and GCA, requiring expert evaluation and leading to delays. This study presents an AI-based diagnostic framework utilizing deep learning models to predict visual scores and classify dementia stages using brain MRI and clinical measures such as TMSE and MoCA. Unlike prior methods that demand full expert oversight, our approach reduces reliance on specialized personnel by enabling AI-generated visual scores to support general radiologists and internists. ResNet18 was trained separately for MTA, ERICA, and GCA scoring, while DenseNet121 was applied for MRI-based dementia classification. Results indicate that models integrating AI-predicted Visual Scores with clinical data achieved up to 75.24% accuracy, outperforming MRI-only models (63.44%). Notably, the inclusion of MoCA unexpectedly reduced classification accuracy, suggesting potential biases in its application. Feature attribution using SHAP revealed that clinical inputs (MMSE, MoCA, age) dominated model decisions, while MRI scores played a greater role in AD classification. Age-stratified confusion matrices further uncovered forward-shifted misclassifications in younger patients, potentially indicating early disease sensitivity. By streamlining the diagnostic process and minimizing the need for licensed specialists, the AI system offers a promising tool for early dementia screening, particularly in areas with limited access to neurologists and radiologists, such as rural Thailand. Future studies will focus on refining model generalizability across diverse populations and improving prediction robustness in real-world clinical settings.

Apical periodontitis (AP) diagnosis primarily relies on periapical radiographs (PRs) and the Periapical Index (PAI) scoring system. However, existing automated approaches often simplify PAI into binary categories or ignore essential clinical metadata, limiting diagnostic performance and applicability. Such limitations hinder timely and accurate diagnosis of AP, which may complicate treatment planning by creating uncertainty about the appropriate timing and type of intervention, and ultimately challenge clinicians' ability to make consistent and informed decisions. This study aimed to develop and validate a novel Mamba-based classification model that integrates PR with structured clinical metadata to predict detailed PAI scores across the full 5-class. In this retrospective diagnostic accuracy study, PRs and corresponding metadata-including patient age, tooth location, tooth number and arch type-were collected from a single institution. Two expert endodontists independently assigned PAI scores (1-5) based on Ørstavik's criteria, with the final reference standard set by consensus. The proposed artificial intelligence (AI) model utilized a Mamba-based state-space architecture to capture spatial dependencies and incorporate structured clinical metadata features. Training and evaluation were conducted using stratified 5-fold cross-validation. The model achieved 54.72% accuracy and a quadratic-weighted kappa (QWK) of 0.713 in 5-class classification, outperforming the latest models based on convolutional neural networks (CNNs) and object detection networks. Ablation analysis further supported the value of integrating patient information, showing that age was the largest impact on model performance. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis for model explainability demonstrated that the model's highlighted areas were aligned with clinically meaningful periapical regions. The proposed model addresses limitations of prior methods by leveraging the full range of the PAI scores and incorporating structured clinical information. It has the potential to support more consistent radiographic interpretation, reduce inter-examiner variability and serve as an interpretable tool in educational and clinical decision-support.

Ambulatory blood pressure monitoring (ABPM) plays an irreplaceable role in the diagnosis and management of hypertension. However, more than 100 million of the 1.4 billion people with hypertension worldwide cannot tolerate nighttime ABPM due to noise and arm compression. Previous prediction methods relying on demographic factors and home blood pressure measurements are time-consuming and burdensome while exhibiting limited accuracy for nocturnal hypertension. There is a need for a more accurate, low-burden approach to identify high-risk patients intolerant to nighttime ABPM monitoring. We collected 2,874 ABPM records at a regional medical center to conduct a retrospective cohort study. Kernel density estimation based preprocessing was applied to stabilize data fluctuations. A variational autoencoder based deep learning model was developed using daytime blood pressure and heart rate combined with full-day activity and posture states to predict nocturnal hypertension. Here we show that the ABPM-VAE model achieves an AUC of 0.82 (95% CI 0.77-0.88) on the test set, outperforming the ablation model (AUC 0.67; 95% CI 0.61-0.74; p<0.001) and prior methods based on demographic and home blood pressure data (AUC 0.69). For nocturnal hypertension prediction, the model yields a PPV of 92.12%, NPV of 55.20%, sensitivity of 0.73, and specificity of 0.84. The entropy reduction preprocessing-enhanced deep learning model predicts nocturnal hypertension risk from ABPM without adding burden to patients or physicians. It serves as an effective screening tool to identify high-risk individuals intolerant to nighttime monitoring, serving as a valuable complement to conventional ABPM.

To the best of our knowledge, this paper is the first to integrate fractal signal processing with vision graph neural networks, establishing a new graph representation learning paradigm consistent with fractal dynamics. Building on this foundation, we propose a Fractal-domain Vision Graph Neural Network (FD-ViG). Specifically, FD-ViG includes: (i) a Fractal-Domain Learning Module that maps images into the fractal-domain using local Hölder exponents and the Singularity Power Spectrum (SPS), enabling fractal-spatial feature fusion; (ii) a Fractal Graph Construction Module that adaptively generates a topology by combining semantic attention with fractal similarity in the fractal feature space; and (iii) a Graph Propagation Module with power-law multi-scale propagation to realize cross-scale diffusion and aggregation, enabling coupled texture-structure learning. Experiments on UCMerced, RSSCN7, and SIRI-WHU achieve overall accuracies of 91.75%, 89.52%, and 92.78%, respectively. Compared with representative vision graph models such as ViG, WiGNet, and ViHGNN, our method achieves consistent improvements over prior methods across all three datasets, while remaining lightweight (2.6M parameters). Moreover, despite having far fewer parameters than ResNet-18, our model yields competitive or better performance on two datasets, and further demonstrates strong generalization ability in cross-dataset evaluation on SAR imagery. This work provides a principled and effective bridge between fractal theory and graph deep learning, benefiting interpretable remote sensing scene understanding under complex textures and structures.

Functional microvascular networks in engineered tissues depend on coordinated endothelial-stromal interactions and evolving extracellular matrix (ECM) mechanics, yet fluorescent staining precludes longitudinal studies and is incompatible with repeated particle-based microrheology and traction force measurements. We develop a deep-learning virtual labeling approach that recovers nuclear, cytoskeletal, and endothelial fluorescence from label-free images acquired in fibrin scaffolds. Human umbilical vein endothelial cells and lung fibroblasts were co-cultured in three-dimensional (3D) fibrin hydrogels containing 2μm silica microbeads that can be used to probe local matrix mechanics. Paired transmission and confocal fluorescence z-stacks (DAPI, phalloidin, UEA I) were used to train a 3D U-Net to generate virtual nuclei, actin, and vascular channels directly from bead-containing label-free volumes. To match channel-specific morphology, edge- and structure-preserving losses were assigned to phalloidin and UEA I, while a sparsity-aware loss was applied to DAPI, improving reconstruction quality across mean squared error, structural similarity, peak signal-to-noise ratio, and correlation. Virtual phalloidin preserved fibrillar density and orientation, virtual UEA I reproduced vessel continuity and density without false matrix labeling, and virtual DAPI enabled nuclei segmentation with Dice scores and total cell counts indistinguishable from ground truth (GT). Microbeads did not generate spurious signal in any channel, demonstrating robustness to scattering particles. Together, these results show that virtual fluorescent labeling can replace destructive staining for quantifying fibrillar architecture, vessel density, and cell number in bead-laden fibrin constructs, and establish a practical route toward longitudinal studies of coupled microvascular morphogenesis and ECM mechanics. Statement of Significance Fluorescent staining is widely used to study vascular growth dynamics in biomaterials but requires destructive fixation, preventing repeated measurements in the same sample. We introduce a label-free deep learning approach that predicts multiple fluorescence markers in fully three-dimensional fibrin hydrogels undergoing capillary morphogenesis. Unlike prior virtual staining methods focused on 2D samples, our model operates in thick, optically complex 3D co-cultures and is robust to highly scattering micron-scale tracer beads used for particle-based micromechanical measurements. A channel-specific loss design preserves sharp boundaries and filament continuity, enabling accurate vascular morphometrics and cell-type identifications that are directly interchangeable with measurements from destructive staining. This approach supports non-destructive, quantitative analysis in tissue-engineering experiments and is extensible to other 3D extracellular matrices.

Gastric cancer (GC) ranks among the most prevalent and lethal cancers worldwide. The albumin-bilirubin (ALBI) score was initially developed to assess liver function in patients with hepatocellular carcinoma. Its modified version, the modified ALBI (mALBI) score, may offer prognostic insights for other malignancies. This study aims to evaluate the prognostic significance of mALBI grade in patients with metastatic GC.Between August 2021 and June 2025, a total of 93 patients with metastatic GC were included in this retrospective multicenter cohort study. mALBI scores were calculated based on serum albumin and total bilirubin concentrations. Patients were classified according to the original mALBI grading system into Grade 1, Grade 2a, Grade 2b, and Grade 3. Overall survival (OS) and progression-free survival (PFS) were assessed using the Kaplan-Meier method, while independent prognostic factors were analyzed via Cox regression.Kaplan-Meier analysis indicated significant differences in OS (χ² = 44.156, p< 0.001) and PFS (χ² = 22.142, p< 0.001) across the four mALBI grades. Utilizing established mALBI grade thresholds and prior binary grouping methods in GC, patients were further divided into low mALBI (Grades 1 and 2a) and high mALBI (Grades 2b and 3) groups. Those with high mALBI grades exhibited significantly worse median OS compared to patients with low mALBI grades (271 vs. 561 days, p= 0.002), as well as significantly shorter median PFS (199 vs. 423 days, p< 0.001). In multivariate analysis, mALBI grade remained independently associated with both OS (p< 0.001) and PFS (p= 0.001).Patients with lower mALBI grades demonstrated considerably better survival outcomes than those with higher grades. Due to its straightforward calculation from routine laboratory tests, mALBI grade may serve as a valuable prognostic marker for survival stratification in metastatic GC. Nonetheless, prospective validation in larger and more diverse cohorts is necessary before its integration into standard clinical practice. Furthermore, additional studies should explore whether interventions aimed at enhancing nutritional status and liver function can positively impact outcomes in high-risk patients.

Semantic clustering under resource constraints via Bayesian low-rank adaptation.

COPD remains a prevalent and debilitating respiratory condition, necessitating early and accurate diagnosis for optimal clinical intervention. In this study, we propose a novel deep learning-based diagnostic framework that employs the ECAPA-TDNN (Emphasized Channel Attention, Propagation and Aggregation—Time Delay Neural Network) architecture to classify respiratory sound signals from the ICBHI dataset. Originally designed for speaker verification, ECAPA-TDNN introduces channel attention and multi-scale feature aggregation, which we adapt for the first time to the domain of medical acoustic analysis. This architecture allows the model to effectively capture subtle and discriminative patterns in pathological breathing sounds, overcoming the limitations of conventional CNN-based methods. Our methodology integrates rigorous signal preprocessing, log-Mel spectrogram extraction, and data augmentation to enhance robustness and generalization. An Attentive Statistics Pooling mechanism is employed for temporal feature summarization, while Grad-CAM-based explainability is incorporated to improve the interpretability of the diagnostic predictions. The model is rigorously validated using a five-fold cross-validation scheme, achieving a mean validation accuracy of 96.8% with consistently high F1-scores and recall rates across all folds. Comparative analysis with prior methods highlights the superiority of our ECAPA-TDNN-based approach in terms of diagnostic precision, robustness, and potential clinical applicability. To the best of our knowledge, this is the first work to adapt ECAPA-TDNN for COPD detection from respiratory sounds, establishing a new benchmark in interpretable and high-performance acoustic-based respiratory disease screening.

Abstract Abstract — The automatic generation of visual scenes from natural language descriptions represents a frontier challenge in artificial intelligence, bridging the gap between linguistic understanding and photorealistic visual synthesis. This paper presents a comprehensive two-phase framework for Text Scene Synthesis. In Phase 1, we leverage a state-of-the-art denoising diffusion probabilistic model (DDPM) conditioned on text embeddings to produce high-fidelity, semantically accurate images from arbitrary textual prompts. In Phase 2, we extend the framework to temporal visual synthesis by employing Contrastive Language-Image Pretraining (CLIP) alongside a video generation pipeline to produce coherent short-form video sequences directly from text. Experimental evaluation demonstrates that the proposed system significantly outperforms prior GAN-based, VAE-based, and standalone diffusion approaches in terms of image fidelity, semantic consistency, and temporal coherence. The system achieves a Fréchet Inception Distance (FID) of 12.4 on COCO-captions and a CLIPSIM score of 0.312 for video generation, representing meaningful improvements over prior methods. This work demonstrates the promise of unified text-to-scene pipelines for applications spanning entertainment, education, game development, and assistive technologies. Keywords — text-to-image synthesis, diffusion models, CLIP, video generation, scene synthesis, generative AI, denoising diffusion probabilistic model, natural language processing.

Multilingual learning is key in natural language processing, but is challenged by the transfer–interference trade-off, where positive transfer benefits certain languages, while negative interference affects others. Prior methods, including linguistic-based and embedding-based language clustering, have attempted to address this; yet, they remain constrained by their static design and lack of task-specific feedback. In this study, we propose a novel computational strategy inspired by molecular design that constructs molecules with targeted properties. Languages are modeled as nodes in an undirected graph, with edges representing the transfer strength. This language molecule is optimized via Reinforcement Learning to adjust edge connections and weights to enhance positive transfer and minimize interference, where graph clustering is applied, and clusters are then evaluated on the Named Entity Recognition and POS tagging tasks using 25 languages from the WikiANN dataset and 12 typologically diverse languages from the UDPOS dataset. Compared to linguistic and embedding-based language clustering baselines, our method yields substantial improvements, especially for low-resource languages, with some showing over 35% increase in F1 score, while high-resource languages benefit moderately, confirming reduced transfer–interference trade-off. Our atom–language model offers a novel path for multilingual learning, inspired by molecular principles from physical sciences.

Brain-Computer Interface (BCI) technology holds great promise for enhancing human health and quality of life, with visual stimulus reconstruction from EEG signals being a key application. However, the complexity and noise of EEG data challenge existing reconstruction methods. To address these issues, we propose NeuroDecoder, an end-to-end multimodal guidance generation framework that produces high-quality images from EEG signals. The key innovation is the collaborative mitigation of EEG noise and cross-modal representation discrepancies through a noise-robust encoder, mask-based triple-contrastive alignment, and a fixed generative model. Specifically, NeuroDecoder consists of three integrated learning stages: 1) EEG Decoding, 2) Modality Alignment, and 3) Image Reconstruction. In the decoding stage, a novel visual decoding model extracts visually relevant features with superior classification accuracy. In the alignment stage, a mask-based triple contrastive learning strategy achieves efficient cross-modal alignment of EEG, text, image, and edge map embeddings into a unified space. In the generation stage, a new reconstruction pipeline feeds the aligned EEG embeddings into a pre-trained stable diffusion model, enabling high-quality visual stimulus reconstruction with enhanced semantic and structural fidelity, without fine-tuning the generative model. On three EEG datasets, NeuroDecoder achieved subject-dependent classification accuracies of 99.76%, 94.41%, and 56.67%, respectively; in the subject-independent setting, it performed near random on EEGCVPR40 but reached 91.61% and 37.63% on the other two. For image reconstruction, it obtained Fréchet Inception Distance of 62.84 and 63.12 on the first two datasets. Extensive experiments demonstrate that NeuroDecoder outperforms prior methods in both EEG classification accuracy and image reconstruction quality.

Epileptic seizure detection from Electroencephalography (EEG) signals is challenging due to their complex temporal dynamics and intricate inter-channel dependencies. A key difficulty lies in capturing multi-scale temporal features that span different time ranges. Seizure-related EEG patterns include rapidly varying micro-scale features and longer-duration macro-scale features, and modeling them jointly often leads to feature interference, hindering accurate temporal representation. In addition, many existing approaches fail to effectively capture spatial relationships between brain regions, further limiting detection performance. To cope with issues above, we design the Adaptive Multi-scale Spatiotemporal Mixing Network (AMSMN). The framework first decomposes EEG signals into macro- and micro-scale sequences, which are processed independently across multiple temporal resolutions to reduce cross-scale interference and better represent temporal dynamics. A spatial attention mechanism then fuses the decomposed features, ensuring that important inter-channel information is preserved. Finally, an Informer-based sparse attention layer captures long-range dependencies, allowing the model to focus on the most relevant global interactions across brain regions. We carry out experiments on two available databases demonstrate that AMSMN consistently achieves superior performance to prior methods in both patient-specific and cross-patient settings. The results confirm that the proposed framework improves seizure detection accuracy, robustness, and generalization by effectively integrating multi-scale temporal modeling with global spatial dependency extraction. This work advances EEG-based seizure detection by enabling precise multi-scale temporal analysis and efficient global dependency modeling, offering strong performance and generalizability for clinical applications.

Digital breast tomosynthesis (DBT) entered the clinical arena in the early 2010s, and initial studies reported encouraging outcomes such as higher cancer detection rates and reduced recall rates. It is important to recognize that when new technologies like DBT are introduced as screening tools, however, their first (prevalent) screens typically detect more cancers than prior methods, known as a prevalence effect. Thus, assessment of multiple rounds of re-screening is necessary to further define DBT's role in population screening and its impact on interval cancer rates and detection of advanced cancers. This article aims to describe what we are learning about DBT beyond the prevalent screening round.