Selected work

Representative Papers

Diffusion MRI with machine learning

Jan 01, 2024Imaging Neuroscience

dMRI analysis faces core challenges including severe noise, high inter-scanner and inter-subject variability, and complex microstructural modeling. This study systematically reviews and empirically evaluates machine learning across the full dMRI pipeline—signal denoising, harmonization, microstructural mapping, fiber tractography, and white-matter pathway quantification—and establishes, for the first time, its applicability boundaries. We innovatively integrate CNNs, GANs, VAEs, transfer learning, multi-site harmonization, and explainable AI (XAI), while proposing a benchmark dataset construction and validation framework tailored for clinical deployment. Our analysis identifies shared bottlenecks in model robustness, reproducibility, and interpretability, and identifies generalizability and standardized evaluation as critical leverage points. The work provides a methodological guide and research roadmap for developing trustworthy, reproducible, and interpretable next-generation dMRI analysis tools.

14 citationsRead paper

Rethinking Medical Landmark Localization with Prototype Learning-based Progressive Offset Correction

Aug 10, 2026

This work addresses the limitations of existing medical image landmark localization methods, which often incur high computational costs in multi-stage optimization and struggle to balance accuracy with efficiency, particularly in anatomically similar regions. To overcome these challenges, the authors propose the PPOC-LL model, which constructs a multi-scale dynamic-aware feature pyramid and integrates a similarity-driven prototype learning mechanism for offset correction. Additionally, an error-aware reliability regularization is introduced to enhance training stability. The proposed method substantially reduces model parameters while achieving a favorable trade-off between high localization accuracy and low computational complexity across multiple public and private X-ray and ultrasound datasets, thereby significantly improving both robustness and efficiency.

0 citationsRead paper

A foundation-model approach to pediatric headache classification from rs-fMRI

Aug 07, 2026

This study addresses the challenge of objectively diagnosing pediatric headache and accurately differentiating its subtypes, which has long been hindered by the lack of reliable biomarkers. For the first time, the authors apply the neuroimaging foundation model NeuroSTORM to pediatric headache classification using resting-state functional MRI (rs-fMRI) data. By leveraging representation learning and fine-tuning, the model achieves effective few-shot transfer without relying on conventional functional connectivity features. In binary classification between headache patients and healthy controls, the model attains an AUROC of 0.82 and an AUPRC of 0.93. Furthermore, in distinguishing among three headache subtypes, it achieves a macro-AUROC of 0.69, significantly outperforming traditional approaches. These results demonstrate the promising potential of foundation models in diagnosing pediatric neurological disorders.

0 citationsRead paper

Recurrent Contrastive Learning for Imbalanced Medical Image Classification

Aug 04, 2026

This work addresses the challenge of class imbalance in medical image classification, where tail classes often exhibit compact representations that are easily dominated by head classes, leading to biased decision boundaries. To mitigate this issue, the authors propose a recursive contrastive learning framework that leverages a temporal memory queue and a temporal anchor mechanism to construct an anchor field for tail classes. By iteratively reusing historical feature states during training, the method progressively expands the latent support region of tail classes, thereby enhancing inter-class separability. Built upon the DINOv3 backbone with LoRA adapters for feature extraction, the approach demonstrates significant performance gains over strong baselines across three imbalanced medical imaging datasets, confirming its effectiveness and robustness.

0 citationsRead paper

DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation

Jul 21, 2026

This work addresses the challenge in 3D medical image segmentation of simultaneously achieving effective global context modeling and parameter efficiency, compounded by the absence of data-driven dynamic scanning mechanisms. To this end, we propose DAMamba, a hybrid U-Net architecture that integrates Mamba state-space modules with triplanar dynamic adaptive scanning (3D-DAS) exclusively into the encoder while retaining convolutional operations elsewhere, enabling efficient fusion of global and local features. The method introduces an encoder-specific DAS embedding strategy that substantially reduces model parameters while enhancing segmentation accuracy. Evaluated on BraTS 2020 using five-fold cross-validation, our base model (5.3M parameters) achieves a Dice score of 0.815, and the larger variant, DAMamba-L (70M parameters), reaches 0.829—outperforming SegMamba with a 13-fold improvement in parameter efficiency.

0 citationsRead paper

Recent publications

Latest Papers

Rethinking Medical Landmark Localization with Prototype Learning-based Progressive Offset Correction

Aug 10, 2026

This work addresses the limitations of existing medical image landmark localization methods, which often incur high computational costs in multi-stage optimization and struggle to balance accuracy with efficiency, particularly in anatomically similar regions. To overcome these challenges, the authors propose the PPOC-LL model, which constructs a multi-scale dynamic-aware feature pyramid and integrates a similarity-driven prototype learning mechanism for offset correction. Additionally, an error-aware reliability regularization is introduced to enhance training stability. The proposed method substantially reduces model parameters while achieving a favorable trade-off between high localization accuracy and low computational complexity across multiple public and private X-ray and ultrasound datasets, thereby significantly improving both robustness and efficiency.

0 citationsRead paper

A foundation-model approach to pediatric headache classification from rs-fMRI

Aug 07, 2026

This study addresses the challenge of objectively diagnosing pediatric headache and accurately differentiating its subtypes, which has long been hindered by the lack of reliable biomarkers. For the first time, the authors apply the neuroimaging foundation model NeuroSTORM to pediatric headache classification using resting-state functional MRI (rs-fMRI) data. By leveraging representation learning and fine-tuning, the model achieves effective few-shot transfer without relying on conventional functional connectivity features. In binary classification between headache patients and healthy controls, the model attains an AUROC of 0.82 and an AUPRC of 0.93. Furthermore, in distinguishing among three headache subtypes, it achieves a macro-AUROC of 0.69, significantly outperforming traditional approaches. These results demonstrate the promising potential of foundation models in diagnosing pediatric neurological disorders.

0 citationsRead paper

Recurrent Contrastive Learning for Imbalanced Medical Image Classification

Aug 04, 2026

This work addresses the challenge of class imbalance in medical image classification, where tail classes often exhibit compact representations that are easily dominated by head classes, leading to biased decision boundaries. To mitigate this issue, the authors propose a recursive contrastive learning framework that leverages a temporal memory queue and a temporal anchor mechanism to construct an anchor field for tail classes. By iteratively reusing historical feature states during training, the method progressively expands the latent support region of tail classes, thereby enhancing inter-class separability. Built upon the DINOv3 backbone with LoRA adapters for feature extraction, the approach demonstrates significant performance gains over strong baselines across three imbalanced medical imaging datasets, confirming its effectiveness and robustness.

0 citationsRead paper

DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation

Jul 21, 2026

This work addresses the challenge in 3D medical image segmentation of simultaneously achieving effective global context modeling and parameter efficiency, compounded by the absence of data-driven dynamic scanning mechanisms. To this end, we propose DAMamba, a hybrid U-Net architecture that integrates Mamba state-space modules with triplanar dynamic adaptive scanning (3D-DAS) exclusively into the encoder while retaining convolutional operations elsewhere, enabling efficient fusion of global and local features. The method introduces an encoder-specific DAS embedding strategy that substantially reduces model parameters while enhancing segmentation accuracy. Evaluated on BraTS 2020 using five-fold cross-validation, our base model (5.3M parameters) achieves a Dice score of 0.815, and the larger variant, DAMamba-L (70M parameters), reaches 0.829—outperforming SegMamba with a 13-fold improvement in parameter efficiency.

0 citationsRead paper

ClinRAG-GRAPH: Clinical-prior Retrieval-Augmented Graph Model with Domain Adversarial Learning for Breast pCR Prediction

Jul 01, 2026

This study addresses key challenges in predicting pathological complete response (pCR) to neoadjuvant chemotherapy in breast cancer—namely, inadequate cross-modal modeling, high inter-center heterogeneity in imaging protocols, and limited interpretability—by integrating dynamic contrast-enhanced MRI (DCE-MRI), clinical variables, and pathological biomarkers. The authors construct intra-patient clinical prior graphs and propose a prior-guided relation-aware graph convolutional network for multimodal representation learning. To mitigate MRI protocol discrepancies and enhance cross-center robustness, a dual-branch domain adversarial strategy is introduced. Furthermore, they pioneer a large language model–driven subgraph retrieval-augmented generation (RAG) mechanism to fuse analogous case evidence and improve model interpretability. The proposed model achieves AUCs of 0.815 on an internal test set and 0.774 and 0.712 on two external test sets, demonstrating strong multicenter generalizability.

0 citationsRead paper