Selected work

Representative Papers

LSEAD: A Privacy-Preserving LLM-Based Speech Analysis Framework for Early Alzheimer's Disease Screening

Aug 07, 2026

This study addresses the critical gap in non-invasive, low-cost methods for early Alzheimer’s disease (AD) screening across diverse clinical settings. The authors propose a privacy-preserving approach that locally deploys open-source large language models to extract embeddings from automatically transcribed speech, followed by dimensionality reduction via principal component analysis (PCA) and machine learning–based classification—all without uploading sensitive patient data. This work represents the first integration of locally hosted large language models with transcribed speech for AD detection, achieving up to a 5% improvement in accuracy on the ADReSS20 and ADReSSo2021 datasets. Notably, the method demonstrates superior performance in the early stages of the disease and exhibits enhanced cross-dataset generalizability.

0 citationsRead paper

What Makes Deep Learning Work for Traditional Chinese Medicine Tongue Diagnosis? A Comprehensive Ablation Study

Jul 30, 2026

This study addresses the lack of clear design principles for multi-label imbalanced medical image classification in automated tongue diagnosis in Traditional Chinese Medicine. Through systematic ablation experiments on two large-scale tongue image datasets (976 and 11,101 samples), it evaluates combinations of six backbone networks, four loss functions, five augmentation strategies, and six training protocols. The work establishes six generalizable design principles: ConvNeXt-Tiny offers superior parameter efficiency; binary cross-entropy loss outperforms asymmetric loss; color-based augmentations should be used sparingly; weak group ensembling surpasses probability averaging; larger data scale substantially boosts performance; and expanding label dimensions risks performance collapse. Under rigorous five-fold cross-validation, the models achieve weighted F1 scores of 0.6625 and 0.7761, respectively, providing reproducible best practices for multi-label imbalanced medical image classification.

0 citationsRead paper

TCMIIES: A Browser-Based LLM-Powered Intelligent Information Extraction System for Academic Literature

May 08, 2026

This work addresses the urgent need among researchers for an accessible, no-code, and privacy-preserving tool to extract structured information from the rapidly growing body of academic literature. The authors propose a browser-based, zero-installation platform that leverages commercial large language model APIs through a schema-guided prompting framework enhanced with automated system prompt generation. Users can define custom extraction schemas via an intuitive graphical interface, while a purely client-side architecture ensures data privacy by keeping all processing within the browser. The system innovatively integrates intelligent field mapping, concurrent batch processing, and automatic retry mechanisms. Evaluated in a traditional Chinese medicine research context, it achieves over 94% compliance in structured output formatting and demonstrates extraction accuracy comparable to that of domain experts for both Chinese and English academic texts.

0 citationsRead paper

Context-Aware Network Based on Multi-scale Spatio-temporal Attention for Action Recognition in Videos

Dec 21, 2025

Existing video action recognition methods neglect the multi-granularity nature of human actions, hindering effective integration of cross-scale spatiotemporal cues. To address this, we propose a Context-Aware Network (CAN) that jointly models temporal dynamics and spatial semantics via two novel modules: a Multi-scale Temporal Cue Module (MTCM) and a Grouped Spatial Cue Module (GSCM). CAN achieves decoupled, adaptive representation by integrating multi-scale spatiotemporal attention, grouped feature-map convolutions, and hierarchical temporal pooling—all enabling end-to-end differentiable training. Evaluated on Something-Something V1/V2, Diving48, Kinetics-400, and UCF101, CAN achieves state-of-the-art accuracy of 50.4%, 63.9%, 88.4%, 74.9%, and 86.9%, respectively. Our approach significantly advances multi-granularity action modeling by explicitly capturing both fine-grained temporal rhythms and hierarchical spatial semantics—from local parts to global configurations—outperforming prevailing methods across diverse benchmarks.

0 citationsRead paper

CEM-FBGTinyDet: Context-Enhanced Foreground Balance with Gradient Tuning for tiny Objects

Jun 11, 2025

In tiny object detection, standard label assignment causes positive sample deficiency in high-level feature pyramid layers (P5–P6), leading to gradient stagnation and semantic degradation; meanwhile, low-level features lack high-level semantic guidance, impairing classification robustness. To address this, we propose E-FPN-BS—a novel framework introducing higher-order feature reuse. It comprises a Context Enhancement Module (CEM) for cross-scale semantic activation, a Foreground-Background Separation Module (FBSM) to improve localization discriminability, and a Scale-Aware Dynamic Gradient Balancing Loss (DCLoss) ensuring equitable gradient backpropagation. Leveraging multi-branch alignment, spatially gated masking, and adaptive weighted fusion, E-FPN-BS achieves consistent mAP gains of +3.2–5.8 across multiple benchmarks, while maintaining strong generalization and real-time deployability.

0 citationsRead paper

Recent publications

Latest Papers

LSEAD: A Privacy-Preserving LLM-Based Speech Analysis Framework for Early Alzheimer's Disease Screening

Aug 07, 2026

This study addresses the critical gap in non-invasive, low-cost methods for early Alzheimer’s disease (AD) screening across diverse clinical settings. The authors propose a privacy-preserving approach that locally deploys open-source large language models to extract embeddings from automatically transcribed speech, followed by dimensionality reduction via principal component analysis (PCA) and machine learning–based classification—all without uploading sensitive patient data. This work represents the first integration of locally hosted large language models with transcribed speech for AD detection, achieving up to a 5% improvement in accuracy on the ADReSS20 and ADReSSo2021 datasets. Notably, the method demonstrates superior performance in the early stages of the disease and exhibits enhanced cross-dataset generalizability.

0 citationsRead paper

What Makes Deep Learning Work for Traditional Chinese Medicine Tongue Diagnosis? A Comprehensive Ablation Study

Jul 30, 2026

This study addresses the lack of clear design principles for multi-label imbalanced medical image classification in automated tongue diagnosis in Traditional Chinese Medicine. Through systematic ablation experiments on two large-scale tongue image datasets (976 and 11,101 samples), it evaluates combinations of six backbone networks, four loss functions, five augmentation strategies, and six training protocols. The work establishes six generalizable design principles: ConvNeXt-Tiny offers superior parameter efficiency; binary cross-entropy loss outperforms asymmetric loss; color-based augmentations should be used sparingly; weak group ensembling surpasses probability averaging; larger data scale substantially boosts performance; and expanding label dimensions risks performance collapse. Under rigorous five-fold cross-validation, the models achieve weighted F1 scores of 0.6625 and 0.7761, respectively, providing reproducible best practices for multi-label imbalanced medical image classification.

0 citationsRead paper

TCMIIES: A Browser-Based LLM-Powered Intelligent Information Extraction System for Academic Literature

May 08, 2026

This work addresses the urgent need among researchers for an accessible, no-code, and privacy-preserving tool to extract structured information from the rapidly growing body of academic literature. The authors propose a browser-based, zero-installation platform that leverages commercial large language model APIs through a schema-guided prompting framework enhanced with automated system prompt generation. Users can define custom extraction schemas via an intuitive graphical interface, while a purely client-side architecture ensures data privacy by keeping all processing within the browser. The system innovatively integrates intelligent field mapping, concurrent batch processing, and automatic retry mechanisms. Evaluated in a traditional Chinese medicine research context, it achieves over 94% compliance in structured output formatting and demonstrates extraction accuracy comparable to that of domain experts for both Chinese and English academic texts.

0 citationsRead paper

Context-Aware Network Based on Multi-scale Spatio-temporal Attention for Action Recognition in Videos

Dec 21, 2025

Existing video action recognition methods neglect the multi-granularity nature of human actions, hindering effective integration of cross-scale spatiotemporal cues. To address this, we propose a Context-Aware Network (CAN) that jointly models temporal dynamics and spatial semantics via two novel modules: a Multi-scale Temporal Cue Module (MTCM) and a Grouped Spatial Cue Module (GSCM). CAN achieves decoupled, adaptive representation by integrating multi-scale spatiotemporal attention, grouped feature-map convolutions, and hierarchical temporal pooling—all enabling end-to-end differentiable training. Evaluated on Something-Something V1/V2, Diving48, Kinetics-400, and UCF101, CAN achieves state-of-the-art accuracy of 50.4%, 63.9%, 88.4%, 74.9%, and 86.9%, respectively. Our approach significantly advances multi-granularity action modeling by explicitly capturing both fine-grained temporal rhythms and hierarchical spatial semantics—from local parts to global configurations—outperforming prevailing methods across diverse benchmarks.

0 citationsRead paper

CEM-FBGTinyDet: Context-Enhanced Foreground Balance with Gradient Tuning for tiny Objects

Jun 11, 2025

In tiny object detection, standard label assignment causes positive sample deficiency in high-level feature pyramid layers (P5–P6), leading to gradient stagnation and semantic degradation; meanwhile, low-level features lack high-level semantic guidance, impairing classification robustness. To address this, we propose E-FPN-BS—a novel framework introducing higher-order feature reuse. It comprises a Context Enhancement Module (CEM) for cross-scale semantic activation, a Foreground-Background Separation Module (FBSM) to improve localization discriminability, and a Scale-Aware Dynamic Gradient Balancing Loss (DCLoss) ensuring equitable gradient backpropagation. Leveraging multi-branch alignment, spatially gated masking, and adaptive weighted fusion, E-FPN-BS achieves consistent mAP gains of +3.2–5.8 across multiple benchmarks, while maintaining strong generalization and real-time deployability.

0 citationsRead paper