Selected work

Representative Papers

Compression-Driven Anomaly Detection in Brain MRI Using an Interpretable Quantum Autoencoder

Jun 25, 2026

This work addresses the challenge of unsupervised anomaly detection in brain MRI, where ground-truth anomaly labels are unavailable, by proposing an interpretable quantum autoencoder-based method. The approach maps image patches to quantum states via angle encoding and employs a variational encoder–decoder architecture augmented with auxiliary “junk” qubits to enable controllable information compression. Anomaly scores are defined based on the incompressibility of inputs relative to normal data. The study highlights the critical role of encoder–decoder asymmetry in detection performance and supports principled threshold selection. Evaluated on public datasets, the model achieves slice-level ROC-AUC of approximately 0.95 and patch-level ROC-AUC of about 0.813, outperforming classical autoencoder and PCA baselines, while producing localized anomaly heatmaps that align well with tumor regions.

0 citationsRead paper

Facial Affect Analysis for Service-Oriented Systems: Advances, Challenges, and Future Visions

Jun 13, 2026

This study reframes facial affect analysis (FAA) from an isolated recognition task into a reusable, composable, and reliable service component within service-oriented software ecosystems (SoSEs). It systematically evaluates the suitability of CNNs, Transformers, graph neural networks, and hybrid architectures for edge–cloud collaborative deployment and, for the first time, introduces SoSE-readiness criteria encompassing uncertainty-aware outputs, runtime assurances, fairness, privacy preservation, and intervention stability. The work establishes a practical FAA service quality attribute framework tailored for real-world deployment, defining measurable interface specifications and lifecycle management mechanisms. This provides an engineering roadmap for integrating FAA capabilities into authentic service ecosystems while ensuring robustness, accountability, and operational viability.

0 citationsRead paper

LLM-Conditioned Synthesis of Pathological Gaits via Structured Gait-Language Representations

Jun 04, 2026

Pathological gait data are extremely scarce due to privacy constraints, difficulties in participant recruitment, high acquisition costs, and substantial inter-individual variability, severely hindering related research. To address this challenge, this work proposes a multimodal large language model–guided generative framework that synthesizes fixed-length 3D skeletal gait sequences from structured textual descriptions. The core innovations include a pathology-aware motion tokenizer that preserves critical pathological motion characteristics in discrete representations, and a semantic enhancement mechanism enabling controllable language-to-gait generation. Under a leave-one-subject-out protocol, a GRU classifier trained on a combination of real and synthetic data achieves an accuracy of 92.77%, demonstrating a significant improvement in downstream classification performance.

0 citationsRead paper

PGcGAN: Pathological Gait-Conditioned GAN for Human Gait Synthesis

Mar 15, 2026

This study addresses the limitations of existing pathological gait analysis, which suffers from small-scale and low-diversity clinical datasets that hinder effective modeling of diverse gait impairments. To overcome this, the authors propose a Pathological Gait conditional Generative Adversarial Network (PGcGAN), which explicitly embeds one-hot encoded pathological labels into the GAN framework for the first time. By integrating a conditional autoencoder architecture, PGcGAN enables controllable synthesis of six distinct gait types through both the generator and discriminator. The model jointly optimizes adversarial and reconstruction objectives, producing highly realistic 3D pose keypoint sequences that preserve structural and temporal characteristics. Experimental results demonstrate that augmenting training data with these synthetic sequences significantly improves the performance of various temporal models—including GRU, LSTM, and CNN—on pathological gait recognition tasks, thereby validating the effectiveness of the proposed approach for data augmentation.

0 citationsRead paper

Illicit object detection in X-ray imaging using deep learning techniques: A comparative evaluation

Jul 23, 2025

X-ray prohibited-item detection faces core challenges including severe occlusion, device heterogeneity, scarce annotated data, and inconsistent evaluation protocols. To address these, this work introduces the first systematic benchmarking framework covering six public X-ray datasets and ten mainstream detection architectures—including CNNs, Transformers, and hybrid models—evaluated using multi-dimensional metrics: mAP₅₀, mAP₅₀:₉₅, inference latency, parameter count, and GFLOPS. Our analysis reveals critical generalization bottlenecks in real-world security screening scenarios and uncovers fundamental compute-accuracy trade-offs. Key findings include: lightweight CNNs demonstrate superior practicality on resource-constrained devices, while Transformers exhibit only marginal gains under heavy occlusion; cross-device generalization remains a persistent challenge. All code, pretrained weights, and evaluation results are publicly released, establishing a reproducible benchmark and actionable guidance for the community.

0 citationsRead paper

Recent publications

Latest Papers

Compression-Driven Anomaly Detection in Brain MRI Using an Interpretable Quantum Autoencoder

Jun 25, 2026

This work addresses the challenge of unsupervised anomaly detection in brain MRI, where ground-truth anomaly labels are unavailable, by proposing an interpretable quantum autoencoder-based method. The approach maps image patches to quantum states via angle encoding and employs a variational encoder–decoder architecture augmented with auxiliary “junk” qubits to enable controllable information compression. Anomaly scores are defined based on the incompressibility of inputs relative to normal data. The study highlights the critical role of encoder–decoder asymmetry in detection performance and supports principled threshold selection. Evaluated on public datasets, the model achieves slice-level ROC-AUC of approximately 0.95 and patch-level ROC-AUC of about 0.813, outperforming classical autoencoder and PCA baselines, while producing localized anomaly heatmaps that align well with tumor regions.

0 citationsRead paper

Facial Affect Analysis for Service-Oriented Systems: Advances, Challenges, and Future Visions

Jun 13, 2026

This study reframes facial affect analysis (FAA) from an isolated recognition task into a reusable, composable, and reliable service component within service-oriented software ecosystems (SoSEs). It systematically evaluates the suitability of CNNs, Transformers, graph neural networks, and hybrid architectures for edge–cloud collaborative deployment and, for the first time, introduces SoSE-readiness criteria encompassing uncertainty-aware outputs, runtime assurances, fairness, privacy preservation, and intervention stability. The work establishes a practical FAA service quality attribute framework tailored for real-world deployment, defining measurable interface specifications and lifecycle management mechanisms. This provides an engineering roadmap for integrating FAA capabilities into authentic service ecosystems while ensuring robustness, accountability, and operational viability.

0 citationsRead paper

LLM-Conditioned Synthesis of Pathological Gaits via Structured Gait-Language Representations

Jun 04, 2026

Pathological gait data are extremely scarce due to privacy constraints, difficulties in participant recruitment, high acquisition costs, and substantial inter-individual variability, severely hindering related research. To address this challenge, this work proposes a multimodal large language model–guided generative framework that synthesizes fixed-length 3D skeletal gait sequences from structured textual descriptions. The core innovations include a pathology-aware motion tokenizer that preserves critical pathological motion characteristics in discrete representations, and a semantic enhancement mechanism enabling controllable language-to-gait generation. Under a leave-one-subject-out protocol, a GRU classifier trained on a combination of real and synthetic data achieves an accuracy of 92.77%, demonstrating a significant improvement in downstream classification performance.

0 citationsRead paper

PGcGAN: Pathological Gait-Conditioned GAN for Human Gait Synthesis

Mar 15, 2026

This study addresses the limitations of existing pathological gait analysis, which suffers from small-scale and low-diversity clinical datasets that hinder effective modeling of diverse gait impairments. To overcome this, the authors propose a Pathological Gait conditional Generative Adversarial Network (PGcGAN), which explicitly embeds one-hot encoded pathological labels into the GAN framework for the first time. By integrating a conditional autoencoder architecture, PGcGAN enables controllable synthesis of six distinct gait types through both the generator and discriminator. The model jointly optimizes adversarial and reconstruction objectives, producing highly realistic 3D pose keypoint sequences that preserve structural and temporal characteristics. Experimental results demonstrate that augmenting training data with these synthetic sequences significantly improves the performance of various temporal models—including GRU, LSTM, and CNN—on pathological gait recognition tasks, thereby validating the effectiveness of the proposed approach for data augmentation.

0 citationsRead paper

Illicit object detection in X-ray imaging using deep learning techniques: A comparative evaluation

Jul 23, 2025

X-ray prohibited-item detection faces core challenges including severe occlusion, device heterogeneity, scarce annotated data, and inconsistent evaluation protocols. To address these, this work introduces the first systematic benchmarking framework covering six public X-ray datasets and ten mainstream detection architectures—including CNNs, Transformers, and hybrid models—evaluated using multi-dimensional metrics: mAP₅₀, mAP₅₀:₉₅, inference latency, parameter count, and GFLOPS. Our analysis reveals critical generalization bottlenecks in real-world security screening scenarios and uncovers fundamental compute-accuracy trade-offs. Key findings include: lightweight CNNs demonstrate superior practicality on resource-constrained devices, while Transformers exhibit only marginal gains under heavy occlusion; cross-device generalization remains a persistent challenge. All code, pretrained weights, and evaluation results are publicly released, establishing a reproducible benchmark and actionable guidance for the community.

0 citationsRead paper