Selected work

Representative Papers

Interpretable Aneurysm Classification via 3D Concept Bottleneck Models: Integrating Morphological and Hemodynamic Clinical Features

Mar 08, 2026

This study addresses the lack of clinical interpretability in deep learning models for intracranial aneurysm classification by proposing the first end-to-end 3D concept bottleneck model. The model maps CTA image features to clinically meaningful concepts—such as morphological and hemodynamic attributes—thereby embedding interpretability aligned with neurosurgical principles directly into the architecture. Built upon pretrained 3D ResNet-34 and DenseNet-121 backbones, the framework incorporates a soft concept bottleneck layer, a composite loss function combining focal loss and concept mean squared error, and eight-fold test-time augmentation (TTA). Experimental results demonstrate that ResNet-34 achieves an accuracy of 93.33% ± 4.5%, while DenseNet-121 reaches 91.43% ± 5.8%; under TTA, the model maintains a stable accuracy of 88.31% with an accuracy–generalization gap below 0.04, effectively balancing high performance with clinical transparency.

0 citationsRead paper

TOSHFA: A Mobile VR-Based System for Pose-Guided Exercise Rehabilitation for Low Back Pain

Jan 24, 2026

This study addresses the challenges of insufficient professional supervision and low patient adherence in home-based low back pain rehabilitation by proposing a low-cost, mobile virtual reality (VR) rehabilitation system. The system leverages a standard laptop webcam combined with MediaPipe to estimate user skeletal keypoints in real time, transmitting this data via low-latency UDP to a Cardboard-style VR headset. To enhance engagement and adherence, it incorporates gamification elements such as point-based rewards and streak-based check-ins to guide therapeutic exercises. As the first home rehabilitation solution integrating lightweight mobile VR with real-time pose estimation, the system demonstrated strong user experience and technical feasibility in a pilot study with 20 participants, laying the groundwork for future multi-exercise clinical trials.

0 citationsRead paper

Efficient Hate Speech Detection: A Three-Layer LoRA-Tuned BERTweet Framework

Nov 08, 2025

To address the challenge of balancing real-time efficiency and detection performance for hate speech in resource-constrained environments, this paper proposes a three-tier collaborative framework: (i) rule-based lightweight pre-filtering; (ii) parameter-efficient fine-tuning of BERTweet via LoRA; and (iii) integration of a continual learning mechanism to adapt to evolving data distributions. The framework requires only 1.87M trainable parameters—just 1.37% of full fine-tuning—and completes training within two hours on a single T4 GPU, achieving a macro-F1 score of 0.85—94% of the performance attained by a 14B-parameter large language model. Its key contributions lie in the synergistic co-design across rule-based, model-level, and learning-level components; high accuracy preservation under extreme parameter efficiency; and an end-to-end lightweight pipeline tailored for edge deployment—significantly reducing computational overhead and deployment costs.

0 citationsRead paper

Learning to Borrow Features for Improved Detection of Small Objects in Single-Shot Detectors

Apr 30, 2025

Single-stage detectors struggle with small-object detection due to semantically impoverished shallow features and low-resolution deep features. To address this, we propose a “cross-layer feature borrowing” paradigm: discriminative deep semantics from large, same-category objects are identified via feature matching, then dynamically integrated into shallow feature maps through weighted aggregation and context-aware fusion—enhancing small-object representations without compromising inference speed or breaking the resolution–semantics trade-off. Our method is built upon the SSD framework and supports end-to-end differentiable training. On the COCO benchmark, it achieves a +4.2% improvement in AP for small objects while maintaining real-time inference speed, demonstrating robustness and generalization across complex scenes.

0 citationsRead paper

Recent publications

Latest Papers

Interpretable Aneurysm Classification via 3D Concept Bottleneck Models: Integrating Morphological and Hemodynamic Clinical Features

Mar 08, 2026

This study addresses the lack of clinical interpretability in deep learning models for intracranial aneurysm classification by proposing the first end-to-end 3D concept bottleneck model. The model maps CTA image features to clinically meaningful concepts—such as morphological and hemodynamic attributes—thereby embedding interpretability aligned with neurosurgical principles directly into the architecture. Built upon pretrained 3D ResNet-34 and DenseNet-121 backbones, the framework incorporates a soft concept bottleneck layer, a composite loss function combining focal loss and concept mean squared error, and eight-fold test-time augmentation (TTA). Experimental results demonstrate that ResNet-34 achieves an accuracy of 93.33% ± 4.5%, while DenseNet-121 reaches 91.43% ± 5.8%; under TTA, the model maintains a stable accuracy of 88.31% with an accuracy–generalization gap below 0.04, effectively balancing high performance with clinical transparency.

0 citationsRead paper

TOSHFA: A Mobile VR-Based System for Pose-Guided Exercise Rehabilitation for Low Back Pain

Jan 24, 2026

This study addresses the challenges of insufficient professional supervision and low patient adherence in home-based low back pain rehabilitation by proposing a low-cost, mobile virtual reality (VR) rehabilitation system. The system leverages a standard laptop webcam combined with MediaPipe to estimate user skeletal keypoints in real time, transmitting this data via low-latency UDP to a Cardboard-style VR headset. To enhance engagement and adherence, it incorporates gamification elements such as point-based rewards and streak-based check-ins to guide therapeutic exercises. As the first home rehabilitation solution integrating lightweight mobile VR with real-time pose estimation, the system demonstrated strong user experience and technical feasibility in a pilot study with 20 participants, laying the groundwork for future multi-exercise clinical trials.

0 citationsRead paper

Efficient Hate Speech Detection: A Three-Layer LoRA-Tuned BERTweet Framework

Nov 08, 2025

To address the challenge of balancing real-time efficiency and detection performance for hate speech in resource-constrained environments, this paper proposes a three-tier collaborative framework: (i) rule-based lightweight pre-filtering; (ii) parameter-efficient fine-tuning of BERTweet via LoRA; and (iii) integration of a continual learning mechanism to adapt to evolving data distributions. The framework requires only 1.87M trainable parameters—just 1.37% of full fine-tuning—and completes training within two hours on a single T4 GPU, achieving a macro-F1 score of 0.85—94% of the performance attained by a 14B-parameter large language model. Its key contributions lie in the synergistic co-design across rule-based, model-level, and learning-level components; high accuracy preservation under extreme parameter efficiency; and an end-to-end lightweight pipeline tailored for edge deployment—significantly reducing computational overhead and deployment costs.

0 citationsRead paper

Learning to Borrow Features for Improved Detection of Small Objects in Single-Shot Detectors

Apr 30, 2025

Single-stage detectors struggle with small-object detection due to semantically impoverished shallow features and low-resolution deep features. To address this, we propose a “cross-layer feature borrowing” paradigm: discriminative deep semantics from large, same-category objects are identified via feature matching, then dynamically integrated into shallow feature maps through weighted aggregation and context-aware fusion—enhancing small-object representations without compromising inference speed or breaking the resolution–semantics trade-off. Our method is built upon the SSD framework and supports end-to-end differentiable training. On the COCO benchmark, it achieves a +4.2% improvement in AP for small objects while maintaining real-time inference speed, demonstrating robustness and generalization across complex scenes.

0 citationsRead paper