Selected work

Representative Papers

Topology-Driven Transferability Estimation for 3D Medical Vision Foundation Models

Jul 05, 2026

Existing methods for evaluating transferability in 3D medical image segmentation rely on time-consuming fine-tuning, which struggles to meet the stringent demands for boundary precision and anatomical consistency. This work proposes the first fine-tuning-free, topology-driven framework that aligns sparse features with semantic labels via minimum spanning trees (MSTs). It assesses transferability at dual scales—local boundary token separability (LBTC) and global representation topological divergence (GRTD)—and incorporates a task-adaptive gated fusion mechanism. Theoretically, we prove that the MST leakage rate constitutes a finite-sample lower bound of the Bayes error and reveal that randomly initialized decoders stabilize topological alignment. Evaluated on a large-scale benchmark encompassing 114,000 3D medical images, our method achieves state-of-the-art performance, improving the weighted Kendall metric by 0.36 on average and accelerating evaluation by 56×.

0 citationsRead paper

DeepFAN, a transformer-based deep learning model for human-artificial intelligence collaborative assessment of incidental pulmonary nodules in CT scans: a multi-reader, multi-case trial

Mar 26, 2026

This study addresses the challenge of effectively integrating global and local features in deep learning–based classification of benign and malignant pulmonary nodules, a task further complicated by the lack of clinical validation in existing approaches. To this end, we propose DeepFAN, a Transformer-based model trained on over 10,000 pathologically confirmed cases—the largest such dataset reported to date—and rigorously evaluated through a multicenter, multi-reader clinical trial. DeepFAN synergistically combines global and local imaging features and incorporates explainability analysis to assess feature contributions. The model achieves an internal test AUC of 0.939 and a clinical trial AUC of 0.954. When used as a diagnostic aid, it improves junior radiologists’ average AUC by 10.9%, with significant gains in sensitivity, specificity, and accuracy, and elevates inter-rater agreement from fair to moderate.

0 citationsRead paper

Fake It Right: Injecting Anatomical Logic into Synthetic Supervised Pre-training for Medical Segmentation

Mar 01, 2026

This work proposes an anatomy-informed synthetic supervised pre-training framework that addresses the limitations of existing methods, which rely on generic geometric shapes and fail to capture the morphological complexity, spatial layout, and inter-organ relationships inherent in real anatomical structures, thereby lacking the global structural priors essential for medical imaging. By integrating anatomical logic—such as spatial anchors and organ topology graphs—into the synthetic data generation process, the framework leverages a lightweight repository of realistic anatomical shapes and a structure-aware sequential placement strategy to enhance physiological plausibility. Evaluated on the Vision Transformer architecture, the method outperforms the current state-of-the-art FDSL baseline by 1.74% on BTCV and surpasses SSL approaches by 1.66% on MSD, while demonstrating robust scalability with increasing synthetic data volume.

0 citationsRead paper

The Texture-Shape Dilemma: Boundary-Safe Synthetic Generation for 3D Medical Transformers

Mar 01, 2026

This work addresses a critical limitation in existing synthetic data approaches for medical Vision Transformers (ViTs), which often neglect the intricate tissue textures and noise inherent in real medical images, leading to inaccurate learning of anatomical boundaries. To resolve this, the authors propose a physics-inspired spatially decoupled synthesis framework that, for the first time, identifies and mitigates an optimization conflict—termed boundary aliasing—between high-frequency texture modeling and boundary delineation. By introducing a gradient masking buffer and a core texture injection mechanism, the method enables orthogonal yet synergistic learning of shape and texture. Integrating equation-driven supervised learning, boundary distance fields, and physically informed spectral texture synthesis, the approach outperforms current FDSL and real-data-pretrained self-supervised methods by 1.43% and 1.51% on the BTCV and MSD datasets, respectively, establishing an efficient, annotation-free, and scalable training paradigm for medical ViTs.

0 citationsRead paper

LDRNet: Large Deformation Registration Model for Chest CT Registration

Feb 02, 2026

This work addresses the challenges of large deformations, complex backgrounds, and regional overlap in chest CT image registration—a domain where existing deep learning approaches, primarily designed for brain images, often underperform. To this end, we propose LDRNet, a fast unsupervised deep registration network that employs a coarse-to-fine multi-resolution strategy. LDRNet integrates two novel components: a Refine Block for multi-scale optimization of the deformation field and a Rigid Block that estimates rigid transformations from high-level features. Evaluated on both a private dataset and the public SegTHOR benchmark, LDRNet significantly outperforms state-of-the-art methods—including VoxelMorph, RCN, and LapIRN—achieving superior registration accuracy while maintaining faster inference speed.

0 citationsRead paper

Recent publications

Latest Papers

Topology-Driven Transferability Estimation for 3D Medical Vision Foundation Models

Jul 05, 2026

Existing methods for evaluating transferability in 3D medical image segmentation rely on time-consuming fine-tuning, which struggles to meet the stringent demands for boundary precision and anatomical consistency. This work proposes the first fine-tuning-free, topology-driven framework that aligns sparse features with semantic labels via minimum spanning trees (MSTs). It assesses transferability at dual scales—local boundary token separability (LBTC) and global representation topological divergence (GRTD)—and incorporates a task-adaptive gated fusion mechanism. Theoretically, we prove that the MST leakage rate constitutes a finite-sample lower bound of the Bayes error and reveal that randomly initialized decoders stabilize topological alignment. Evaluated on a large-scale benchmark encompassing 114,000 3D medical images, our method achieves state-of-the-art performance, improving the weighted Kendall metric by 0.36 on average and accelerating evaluation by 56×.

0 citationsRead paper

DeepFAN, a transformer-based deep learning model for human-artificial intelligence collaborative assessment of incidental pulmonary nodules in CT scans: a multi-reader, multi-case trial

Mar 26, 2026

This study addresses the challenge of effectively integrating global and local features in deep learning–based classification of benign and malignant pulmonary nodules, a task further complicated by the lack of clinical validation in existing approaches. To this end, we propose DeepFAN, a Transformer-based model trained on over 10,000 pathologically confirmed cases—the largest such dataset reported to date—and rigorously evaluated through a multicenter, multi-reader clinical trial. DeepFAN synergistically combines global and local imaging features and incorporates explainability analysis to assess feature contributions. The model achieves an internal test AUC of 0.939 and a clinical trial AUC of 0.954. When used as a diagnostic aid, it improves junior radiologists’ average AUC by 10.9%, with significant gains in sensitivity, specificity, and accuracy, and elevates inter-rater agreement from fair to moderate.

0 citationsRead paper

Fake It Right: Injecting Anatomical Logic into Synthetic Supervised Pre-training for Medical Segmentation

Mar 01, 2026

This work proposes an anatomy-informed synthetic supervised pre-training framework that addresses the limitations of existing methods, which rely on generic geometric shapes and fail to capture the morphological complexity, spatial layout, and inter-organ relationships inherent in real anatomical structures, thereby lacking the global structural priors essential for medical imaging. By integrating anatomical logic—such as spatial anchors and organ topology graphs—into the synthetic data generation process, the framework leverages a lightweight repository of realistic anatomical shapes and a structure-aware sequential placement strategy to enhance physiological plausibility. Evaluated on the Vision Transformer architecture, the method outperforms the current state-of-the-art FDSL baseline by 1.74% on BTCV and surpasses SSL approaches by 1.66% on MSD, while demonstrating robust scalability with increasing synthetic data volume.

0 citationsRead paper

The Texture-Shape Dilemma: Boundary-Safe Synthetic Generation for 3D Medical Transformers

Mar 01, 2026

This work addresses a critical limitation in existing synthetic data approaches for medical Vision Transformers (ViTs), which often neglect the intricate tissue textures and noise inherent in real medical images, leading to inaccurate learning of anatomical boundaries. To resolve this, the authors propose a physics-inspired spatially decoupled synthesis framework that, for the first time, identifies and mitigates an optimization conflict—termed boundary aliasing—between high-frequency texture modeling and boundary delineation. By introducing a gradient masking buffer and a core texture injection mechanism, the method enables orthogonal yet synergistic learning of shape and texture. Integrating equation-driven supervised learning, boundary distance fields, and physically informed spectral texture synthesis, the approach outperforms current FDSL and real-data-pretrained self-supervised methods by 1.43% and 1.51% on the BTCV and MSD datasets, respectively, establishing an efficient, annotation-free, and scalable training paradigm for medical ViTs.

0 citationsRead paper

LDRNet: Large Deformation Registration Model for Chest CT Registration

Feb 02, 2026

This work addresses the challenges of large deformations, complex backgrounds, and regional overlap in chest CT image registration—a domain where existing deep learning approaches, primarily designed for brain images, often underperform. To this end, we propose LDRNet, a fast unsupervised deep registration network that employs a coarse-to-fine multi-resolution strategy. LDRNet integrates two novel components: a Refine Block for multi-scale optimization of the deformation field and a Rigid Block that estimates rigid transformations from high-level features. Evaluated on both a private dataset and the public SegTHOR benchmark, LDRNet significantly outperforms state-of-the-art methods—including VoxelMorph, RCN, and LapIRN—achieving superior registration accuracy while maintaining faster inference speed.

0 citationsRead paper