Selected work

Representative Papers

Structural Graph Neural Networks with Anatomical Priors for Explainable Chest X-ray Diagnosis

Jan 17, 2026

This work addresses the challenges of interpretability and precise lesion localization in chest X-ray diagnosis by proposing a graph neural network framework that integrates anatomical structure priors. The method reformulates convolutional feature maps into patch-level graphs incorporating both appearance features and spatial coordinates, and introduces a tailored structural propagation mechanism that explicitly models relative anatomical relationships among nodes, thereby endowing the graph network with an inductive bias for structured reasoning. Intrinsic interpretability is achieved through node importance scoring and joint graph-node prediction, eliminating the need for post-hoc visualization. Experimental results demonstrate that the proposed approach effectively enhances both diagnostic accuracy and model interpretability, while exhibiting strong potential for cross-domain generalization.

0 citationsRead paper

APC-GNN++: An Adaptive Patient-Centric GNN with Context-Aware Attention and Mini-Graph Explainability for Diabetes Classification

Dec 20, 2025

To address challenges in diabetes clinical classification—including difficulty modeling patient relationships, poor generalization to unseen patients, and insufficient prediction interpretability—this paper proposes an Adaptive Patient-Centered Graph Neural Network (PC-GNN). Methodologically, it introduces a context-aware edge attention mechanism to capture clinically meaningful patient associations; incorporates confidence-guided node-graph feature fusion and neighborhood consistency regularization to balance individual specificity and population-level patterns; and pioneers a mini-graph construction strategy enabling real-time, interpretable predictions for unseen patients. Evaluated on a real-world hospital dataset from Algeria, PC-GNN significantly outperforms MLP, Random Forest, XGBoost, and GCN, achieving +3.2% test accuracy and +4.7% macro-F1 score. A companion Tkinter-based GUI system supports node-level confidence analysis and interactive clinical decision-making.

0 citationsRead paper

A Hybrid Deep Learning and Anomaly Detection Framework for Real-Time Malicious URL Classification

Nov 30, 2025

Malicious URLs constitute a primary vector for phishing and malware distribution, necessitating real-time detection methods with low latency and high robustness. This paper proposes a multi-stage classification framework integrating deep learning and anomaly detection. First, lightweight statistical features are extracted via hash-based n-gram vectorization. Second, Isolation Forest is employed for anomaly filtering to enhance robustness against obfuscated URLs. Finally, a lightweight neural network performs binary classification. The system incorporates SMOTE for class imbalance mitigation, a multilingual Tkinter-based GUI, and clipboard-triggered automatic scanning. Evaluated on public benchmark datasets, the model achieves 96.4% accuracy, 95.4% F1-score, and 97.3% ROC-AUC, with an average inference latency of only 20 ms per sample—50–100× faster than CNN- or SVM-based baselines. These advances significantly improve practical deployability and scalability for real-time URL threat detection.

0 citationsRead paper

Recent publications

Latest Papers

Structural Graph Neural Networks with Anatomical Priors for Explainable Chest X-ray Diagnosis

Jan 17, 2026

This work addresses the challenges of interpretability and precise lesion localization in chest X-ray diagnosis by proposing a graph neural network framework that integrates anatomical structure priors. The method reformulates convolutional feature maps into patch-level graphs incorporating both appearance features and spatial coordinates, and introduces a tailored structural propagation mechanism that explicitly models relative anatomical relationships among nodes, thereby endowing the graph network with an inductive bias for structured reasoning. Intrinsic interpretability is achieved through node importance scoring and joint graph-node prediction, eliminating the need for post-hoc visualization. Experimental results demonstrate that the proposed approach effectively enhances both diagnostic accuracy and model interpretability, while exhibiting strong potential for cross-domain generalization.

0 citationsRead paper

APC-GNN++: An Adaptive Patient-Centric GNN with Context-Aware Attention and Mini-Graph Explainability for Diabetes Classification

Dec 20, 2025

To address challenges in diabetes clinical classification—including difficulty modeling patient relationships, poor generalization to unseen patients, and insufficient prediction interpretability—this paper proposes an Adaptive Patient-Centered Graph Neural Network (PC-GNN). Methodologically, it introduces a context-aware edge attention mechanism to capture clinically meaningful patient associations; incorporates confidence-guided node-graph feature fusion and neighborhood consistency regularization to balance individual specificity and population-level patterns; and pioneers a mini-graph construction strategy enabling real-time, interpretable predictions for unseen patients. Evaluated on a real-world hospital dataset from Algeria, PC-GNN significantly outperforms MLP, Random Forest, XGBoost, and GCN, achieving +3.2% test accuracy and +4.7% macro-F1 score. A companion Tkinter-based GUI system supports node-level confidence analysis and interactive clinical decision-making.

0 citationsRead paper

A Hybrid Deep Learning and Anomaly Detection Framework for Real-Time Malicious URL Classification

Nov 30, 2025

Malicious URLs constitute a primary vector for phishing and malware distribution, necessitating real-time detection methods with low latency and high robustness. This paper proposes a multi-stage classification framework integrating deep learning and anomaly detection. First, lightweight statistical features are extracted via hash-based n-gram vectorization. Second, Isolation Forest is employed for anomaly filtering to enhance robustness against obfuscated URLs. Finally, a lightweight neural network performs binary classification. The system incorporates SMOTE for class imbalance mitigation, a multilingual Tkinter-based GUI, and clipboard-triggered automatic scanning. Evaluated on public benchmark datasets, the model achieves 96.4% accuracy, 95.4% F1-score, and 97.3% ROC-AUC, with an average inference latency of only 20 ms per sample—50–100× faster than CNN- or SVM-based baselines. These advances significantly improve practical deployability and scalability for real-time URL threat detection.

0 citationsRead paper