Selected work

Representative Papers

Phenotyping TPF via Self-Supervised Learning: A Label-Agnostic Framework with Expert Validation

Jun 15, 2026

This study addresses the substantial inter-observer variability inherent in conventional tibial plateau fracture classification systems—such as Schatzker and AO/OTA—which can cause supervised models to learn human disagreement rather than true morphological patterns. To overcome this limitation, the authors propose the first fully unsupervised self-supervised learning framework that leverages knee radiographs to automatically discover imaging-derived fracture phenotypes. The approach integrates RadImageNet-pretrained ResNet-50, SimCLR contrastive learning, UMAP dimensionality reduction, and k-means clustering. Clinical expert blind review confirmed that the four identified phenotypes exhibit high cohesion (silhouette coefficient: 0.511; bootstrap-adjusted Rand index [ARI]: 0.319) and strong clinical interpretability. Notably, one phenotype was consistently interpreted as comminuted fracture and proved nearly orthogonal to Schatzker classification (ARI = 0.013), revealing a critical morphological dimension overlooked by existing taxonomies.

0 citationsRead paper

3D-CBM: A Framework for Concept-Based Interpretability in Generative 3D Modeling

Jun 09, 2026

This work addresses the lack of interpretability in 3D generative models for deep geometric learning by introducing Concept Bottleneck Models (CBMs) into unstructured 3D data generation, establishing an intrinsically interpretable framework. The proposed method maps point clouds or meshes onto a hierarchy of human-understandable semantic concepts—such as part-level structures and functional attributes—enabling interactive intervention and semantic manipulation at test time. Experiments on PartNet and ShapeNet demonstrate that the model achieves a part-level concept prediction accuracy of 88.8% and a Chamfer Distance of 0.0115, while effectively correcting structural errors. This approach provides a foundational architecture for human-in-the-loop collaborative design in 3D shape generation.

0 citationsRead paper

PAIRED: A Process-Anchored Framework for Transparent Reporting of AI Contributions in Scientific Research

May 22, 2026

Current AI contribution disclosure mechanisms focus solely on final outputs, failing to accurately capture the cognitive interplay and division of decision-making between humans and AI throughout the research process, thereby leading to ambiguous attribution. This work proposes the PAIRED framework, which introduces a process-oriented disclosure mechanism centered on “decision points” as the fundamental unit, enabling prospective author logs to automatically generate structured contribution records. Grounded in four core design principles—decision-point granularity, dual-sided output, artifact-triggered logging, and process-anchored modeling—the framework is embedded within AI-enabled research platforms to facilitate automated, compliant disclosure. Empirical validation demonstrates that PAIRED effectively distinguishes human originality from the degree of AI suggestion adoption and offers a viable pathway for platform integration, substantially enhancing transparency and traceability in AI-augmented scientific collaboration.

0 citationsRead paper

Synthesizing the Expert: A Validated Multimodal Dataset for Trustworthy AI-Assisted Swimming Coaching

May 12, 2026

This work addresses the challenges of scarce real-world underwater data, ethical constraints on biological data collection, and high costs of expert annotation by proposing a multi-agent large language model–based synthetic data generation framework. Integrating physiological signals, kinematic sensor data, scientific literature, and expert-derived rules, the study constructs the first structured, trustworthy benchmark dataset tailored for AI-powered swimming coaching. The system generates 1,864 high-quality question-context-answer triplets, validated against 12 physiologically plausible rules to ensure fidelity. This rule-driven, multimodal knowledge fusion approach enables reliable retrieval-augmented generation (RAG) and establishes a foundational data resource for high-fidelity, trustworthy AI-assisted swim training systems.

0 citationsRead paper

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

Recent publications

Latest Papers

Phenotyping TPF via Self-Supervised Learning: A Label-Agnostic Framework with Expert Validation

Jun 15, 2026

This study addresses the substantial inter-observer variability inherent in conventional tibial plateau fracture classification systems—such as Schatzker and AO/OTA—which can cause supervised models to learn human disagreement rather than true morphological patterns. To overcome this limitation, the authors propose the first fully unsupervised self-supervised learning framework that leverages knee radiographs to automatically discover imaging-derived fracture phenotypes. The approach integrates RadImageNet-pretrained ResNet-50, SimCLR contrastive learning, UMAP dimensionality reduction, and k-means clustering. Clinical expert blind review confirmed that the four identified phenotypes exhibit high cohesion (silhouette coefficient: 0.511; bootstrap-adjusted Rand index [ARI]: 0.319) and strong clinical interpretability. Notably, one phenotype was consistently interpreted as comminuted fracture and proved nearly orthogonal to Schatzker classification (ARI = 0.013), revealing a critical morphological dimension overlooked by existing taxonomies.

0 citationsRead paper

3D-CBM: A Framework for Concept-Based Interpretability in Generative 3D Modeling

Jun 09, 2026

This work addresses the lack of interpretability in 3D generative models for deep geometric learning by introducing Concept Bottleneck Models (CBMs) into unstructured 3D data generation, establishing an intrinsically interpretable framework. The proposed method maps point clouds or meshes onto a hierarchy of human-understandable semantic concepts—such as part-level structures and functional attributes—enabling interactive intervention and semantic manipulation at test time. Experiments on PartNet and ShapeNet demonstrate that the model achieves a part-level concept prediction accuracy of 88.8% and a Chamfer Distance of 0.0115, while effectively correcting structural errors. This approach provides a foundational architecture for human-in-the-loop collaborative design in 3D shape generation.

0 citationsRead paper

PAIRED: A Process-Anchored Framework for Transparent Reporting of AI Contributions in Scientific Research

May 22, 2026

Current AI contribution disclosure mechanisms focus solely on final outputs, failing to accurately capture the cognitive interplay and division of decision-making between humans and AI throughout the research process, thereby leading to ambiguous attribution. This work proposes the PAIRED framework, which introduces a process-oriented disclosure mechanism centered on “decision points” as the fundamental unit, enabling prospective author logs to automatically generate structured contribution records. Grounded in four core design principles—decision-point granularity, dual-sided output, artifact-triggered logging, and process-anchored modeling—the framework is embedded within AI-enabled research platforms to facilitate automated, compliant disclosure. Empirical validation demonstrates that PAIRED effectively distinguishes human originality from the degree of AI suggestion adoption and offers a viable pathway for platform integration, substantially enhancing transparency and traceability in AI-augmented scientific collaboration.

0 citationsRead paper

Synthesizing the Expert: A Validated Multimodal Dataset for Trustworthy AI-Assisted Swimming Coaching

May 12, 2026

This work addresses the challenges of scarce real-world underwater data, ethical constraints on biological data collection, and high costs of expert annotation by proposing a multi-agent large language model–based synthetic data generation framework. Integrating physiological signals, kinematic sensor data, scientific literature, and expert-derived rules, the study constructs the first structured, trustworthy benchmark dataset tailored for AI-powered swimming coaching. The system generates 1,864 high-quality question-context-answer triplets, validated against 12 physiologically plausible rules to ensure fidelity. This rule-driven, multimodal knowledge fusion approach enables reliable retrieval-augmented generation (RAG) and establishes a foundational data resource for high-fidelity, trustworthy AI-assisted swim training systems.

0 citationsRead paper

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