Projects

An annotation platform for chest CT that integrates DICOM preprocessing, candidate nodule detection, and browser-based image viewing. It organizes review by batch, patient, examination date, and CT series, with independent multi-reviewer assessments, candidate-level decisions, and batch access control. [Code]

Mitigating Entity Hallucinations in 3D Radiology Report Generation via Dual-Stream Alignment. [Code]

Efficient and Gender-Adaptive Graph Vision Mamba for Pediatric Bone Age Assessment. [Code]

Rethinking Foreground Consistency for Efficient Semi-Supervised 3D Medical Image Segmentation. [Code]

Building ChestWorld on 118,591 chest CT scans with patient-level data isolation, multiscale JEPA and CTM modeling, and region-level self-supervision. Training and evaluation use multiple seeds and bootstrap confidence intervals; longitudinal CT prediction and image–report alignment remain planned extensions. [Code]

Developed a ReAct workflow for CT / MRI / X-ray reporting with traceable evidence, tool validation, human review, and offline regression. Curated 28,752 image–report pairs with hospital collaborators and packaged a PACS-compatible inference service.

Built a patient-level risk stratification pipeline combining DICOM quality control, ForeMatch segmentation, radiomics, Graph Mamba representations, and risk calibration. Added robust modeling, layered quality checks, and versioned audit trails, and designed DICOM/PACS integration interfaces.