Selected work

Representative Papers

LeukocyteCount: Automatic Identification and Counting for leukocytes using Deep Learning

Jul 05, 2026

This study addresses the limitations of conventional manual white blood cell (WBC) counting and classification, which suffer from low efficiency and susceptibility to human error, thereby compromising diagnostic accuracy. To overcome these challenges, this work proposes an automated deep learning–based approach that innovatively integrates YOLOv5 for object detection, a fine-tuned MobileNetV2 for feature extraction, and a logistic regression classifier to achieve precise WBC detection, enumeration, and four-class classification. Experimental evaluation on the BCCD dataset demonstrates that the proposed method attains a WBC detection accuracy of 98%, a classification accuracy of 99.04%, and an F1 score of 99.73% for red blood cell detection, significantly outperforming existing baseline methods.

0 citationsRead paper

HieroGlyphTranslator: Automatic Recognition and Translation of Egyptian Hieroglyphs to English

Dec 03, 2025

This paper addresses the end-to-end automatic translation of ancient Egyptian hieroglyphic images into English—a challenging task due to polysemy (one glyph, multiple meanings) and complex spatial layouts. We propose a three-stage framework: (1) robust glyph segmentation leveraging contour detection and Detectron2; (2) standardized semantic mapping using Gardiner codes as an intermediate linguistic representation; and (3) CNN-based sequence-to-sequence translation. Crucially, we embed symbol-level structural priors—namely, the Gardiner classification system—into the translation pipeline to mitigate semantic ambiguity. The method is trained and evaluated on the Morris Franken and EgyptianTranslation datasets, achieving a BLEU score of 42.2—substantially outperforming prior approaches. To our knowledge, this is the first work to achieve both high-accuracy and interpretable direct image-to-text translation of hieroglyphs, enabling faithful, linguistically grounded reconstructions without manual transcription intermediaries.

0 citationsRead paper

Adaptive Dual-Layer Web Application Firewall (ADL-WAF) Leveraging Machine Learning for Enhanced Anomaly and Threat Detection

Nov 16, 2025

Traditional web application firewalls (WAFs) suffer from high false-positive rates due to their inability to accurately distinguish malicious traffic from legitimate requests. To address this, this paper proposes an adaptive two-tier machine learning–based WAF architecture. The architecture employs a cascaded decision tree for anomaly detection followed by a support vector machine for fine-grained threat classification, integrated with comprehensive feature engineering and data preprocessing. It is trained and validated on five publicly available benchmark datasets. Compared to single-model approaches, the collaborative two-tier design significantly enhances discrimination capability: achieving an average detection accuracy of 99.88%, perfect precision (100%), and substantially reduced false positives. The core innovation lies in decoupling and cascading anomaly detection and attack-type classification—enabling both real-time responsiveness and granular threat identification. This framework provides a reproducible, robust, and scalable pathway toward intelligent WAF evolution.

0 citationsRead paper

A Novel Approach to Breast Cancer Segmentation using U-Net Model with Attention Mechanisms and FedProx

Oct 21, 2025

Breast cancer ultrasound image segmentation faces dual challenges: stringent patient privacy requirements and model bias induced by non-independent and identically distributed (non-IID) data across clients. To address these, we propose a federated learning framework integrating FedProx optimization with an attention-enhanced U-Net architecture. FedProx mitigates convergence instability caused by local data heterogeneity, while the attention mechanism strengthens discriminative feature learning—particularly at tumor boundaries—and the U-Net backbone ensures high segmentation fidelity. Our method achieves 96% global segmentation accuracy without requiring raw data to leave local sites, significantly outperforming standard federated baselines. The key contribution is the first synergistic integration of FedProx and attention-augmented U-Net specifically for medical image segmentation in federated settings—thereby jointly ensuring privacy preservation, model robustness under data heterogeneity, and clinically competitive segmentation performance.

0 citationsRead paper

Advancing Brain Tumor Segmentation via Attention-based 3D U-Net Architecture and Digital Image Processing

Oct 21, 2025

Addressing challenges in brain tumor MRI segmentation—including irregular shape delineation, ambiguous boundaries, and severe class imbalance—this paper proposes an enhanced 3D U-Net integrating attention mechanisms with digital image processing techniques. Specifically, a channel-spatial collaborative attention module is introduced to strengthen discriminative feature representation at critical regions, while a pre-detection-guided coarse tumor localization strategy mitigates data imbalance and improves model generalization and robustness. Evaluated on the BraTS 2020 dataset, the model achieves Dice score 0.975, specificity 0.988, and sensitivity 0.995—outperforming state-of-the-art methods. The main contributions are: (1) an attention-driven multi-scale feature focusing mechanism; (2) an image-processing-assisted weakly supervised pre-detection paradigm; and (3) a lightweight, efficient segmentation framework tailored for high-resolution medical imaging.

0 citationsRead paper

Recent publications

Latest Papers

LeukocyteCount: Automatic Identification and Counting for leukocytes using Deep Learning

Jul 05, 2026

This study addresses the limitations of conventional manual white blood cell (WBC) counting and classification, which suffer from low efficiency and susceptibility to human error, thereby compromising diagnostic accuracy. To overcome these challenges, this work proposes an automated deep learning–based approach that innovatively integrates YOLOv5 for object detection, a fine-tuned MobileNetV2 for feature extraction, and a logistic regression classifier to achieve precise WBC detection, enumeration, and four-class classification. Experimental evaluation on the BCCD dataset demonstrates that the proposed method attains a WBC detection accuracy of 98%, a classification accuracy of 99.04%, and an F1 score of 99.73% for red blood cell detection, significantly outperforming existing baseline methods.

0 citationsRead paper

HieroGlyphTranslator: Automatic Recognition and Translation of Egyptian Hieroglyphs to English

Dec 03, 2025

This paper addresses the end-to-end automatic translation of ancient Egyptian hieroglyphic images into English—a challenging task due to polysemy (one glyph, multiple meanings) and complex spatial layouts. We propose a three-stage framework: (1) robust glyph segmentation leveraging contour detection and Detectron2; (2) standardized semantic mapping using Gardiner codes as an intermediate linguistic representation; and (3) CNN-based sequence-to-sequence translation. Crucially, we embed symbol-level structural priors—namely, the Gardiner classification system—into the translation pipeline to mitigate semantic ambiguity. The method is trained and evaluated on the Morris Franken and EgyptianTranslation datasets, achieving a BLEU score of 42.2—substantially outperforming prior approaches. To our knowledge, this is the first work to achieve both high-accuracy and interpretable direct image-to-text translation of hieroglyphs, enabling faithful, linguistically grounded reconstructions without manual transcription intermediaries.

0 citationsRead paper

Adaptive Dual-Layer Web Application Firewall (ADL-WAF) Leveraging Machine Learning for Enhanced Anomaly and Threat Detection

Nov 16, 2025

Traditional web application firewalls (WAFs) suffer from high false-positive rates due to their inability to accurately distinguish malicious traffic from legitimate requests. To address this, this paper proposes an adaptive two-tier machine learning–based WAF architecture. The architecture employs a cascaded decision tree for anomaly detection followed by a support vector machine for fine-grained threat classification, integrated with comprehensive feature engineering and data preprocessing. It is trained and validated on five publicly available benchmark datasets. Compared to single-model approaches, the collaborative two-tier design significantly enhances discrimination capability: achieving an average detection accuracy of 99.88%, perfect precision (100%), and substantially reduced false positives. The core innovation lies in decoupling and cascading anomaly detection and attack-type classification—enabling both real-time responsiveness and granular threat identification. This framework provides a reproducible, robust, and scalable pathway toward intelligent WAF evolution.

0 citationsRead paper

A Novel Approach to Breast Cancer Segmentation using U-Net Model with Attention Mechanisms and FedProx

Oct 21, 2025

Breast cancer ultrasound image segmentation faces dual challenges: stringent patient privacy requirements and model bias induced by non-independent and identically distributed (non-IID) data across clients. To address these, we propose a federated learning framework integrating FedProx optimization with an attention-enhanced U-Net architecture. FedProx mitigates convergence instability caused by local data heterogeneity, while the attention mechanism strengthens discriminative feature learning—particularly at tumor boundaries—and the U-Net backbone ensures high segmentation fidelity. Our method achieves 96% global segmentation accuracy without requiring raw data to leave local sites, significantly outperforming standard federated baselines. The key contribution is the first synergistic integration of FedProx and attention-augmented U-Net specifically for medical image segmentation in federated settings—thereby jointly ensuring privacy preservation, model robustness under data heterogeneity, and clinically competitive segmentation performance.

0 citationsRead paper

Advancing Brain Tumor Segmentation via Attention-based 3D U-Net Architecture and Digital Image Processing

Oct 21, 2025

Addressing challenges in brain tumor MRI segmentation—including irregular shape delineation, ambiguous boundaries, and severe class imbalance—this paper proposes an enhanced 3D U-Net integrating attention mechanisms with digital image processing techniques. Specifically, a channel-spatial collaborative attention module is introduced to strengthen discriminative feature representation at critical regions, while a pre-detection-guided coarse tumor localization strategy mitigates data imbalance and improves model generalization and robustness. Evaluated on the BraTS 2020 dataset, the model achieves Dice score 0.975, specificity 0.988, and sensitivity 0.995—outperforming state-of-the-art methods. The main contributions are: (1) an attention-driven multi-scale feature focusing mechanism; (2) an image-processing-assisted weakly supervised pre-detection paradigm; and (3) a lightweight, efficient segmentation framework tailored for high-resolution medical imaging.

0 citationsRead paper