Vision Foundation Models
01Generalizable visual representations adaptable across imaging modalities and organs for lesion detection, anatomical segmentation, risk stratification, and quantitative biomarkers.
Department of Interventional & Diagnostic Radiology · Institute of Medical Imaging, Shanghai Jiao Tong University
We develop artificial intelligence methods for disease detection, diagnostic reasoning, quantitative analysis, and image reconstruction in clinical radiology — evaluated through multicenter cohorts and radiologist reader studies, while investigating physiological and pathological mechanisms reflected in medical imaging.
Our work examines how computational models represent medical images, integrate clinical context, restore fine anatomical detail, and connect image-derived structures with physiological function.
Generalizable visual representations adaptable across imaging modalities and organs for lesion detection, anatomical segmentation, risk stratification, and quantitative biomarkers.
Combining medical images, radiology reports, clinical history, and diagnostic criteria for structured reporting, multi-agent clinical reasoning, and reader-validated decision support.
Super-resolution, denoising, artifact restoration, and cross-modal synthesis designed to preserve diagnostically critical micro-architecture and quantitative accuracy.
3D anatomical modeling, registration, vascular hemodynamics, and musculoskeletal biomechanics linking structural imaging phenotypes to physiological function.
Large-scale representation learning, criteria-grounded clinical reasoning, and micro-architectural super-resolution validated across multicenter cohorts and structured radiologist reader studies.
Built on BoneFM, a DINOv2-style self-supervised whole-body skeletal CT backbone, BoneCoT encodes clinician-derived task dependencies to guide step-by-step reasoning from anatomical localization and bone density changes to interpretable bone metastasis assessment across multicenter cohorts.
Pretrained via self-supervision on 103,989 non-contrast abdominal CT scans and fine-tuned on 5,816 annotated cases across 11 acute abdominal conditions and 3 risk-stratification tasks. Across 2,528 external patients from 3 cohorts, macro-AUROC reached 0.919 on emergent conditions; in a crossover reader study, AI assistance raised mean AUROC from 0.812 to 0.924 and cut median reading time by 52.5 s per case.
Emulates a multidisciplinary stroke team via specialized agents coordinating emergency triage, time-window estimation, multimodal neuroimaging interpretation, contraindication screening, and retrieval-augmented guideline reasoning. Evaluated on 1,658 suspected acute ischemic stroke cases across two centers, improving treatment recommendation accuracy from 0.636 to 0.785 (Macro-F1 0.562 → 0.768).
Establishes a reproducible multicenter benchmark and blinded radiologist reader study evaluating 10 mainstream large language models on generating diagnostic impressions from free-text brain MRI report findings, systematically quantifying factual accuracy, clinical completeness, and reporting utility.
Releases the first spatially registered paired dataset of routine clinical CT and high-resolution ex-vivo micro-CT volumes of human lumbar vertebrae, bridging in-vivo clinical imaging and trabecular bone micro-architecture through an open super-resolution benchmark and community challenge.
Addresses rapid multi-source clinical synthesis within narrow stroke reperfusion windows by orchestrating specialized agents for structured imaging readout, neurological feature extraction, and guideline-aligned treatment recommendation, validated across multicenter cohorts with an interactive web demo.
Combining 3D vascular geometry, hemodynamic constraints, biomechanical trajectory simulation, and multi-reader clinical studies across vascular, cardiopulmonary, and musculoskeletal disorders.
Integrates 3D vascular point-cloud representation learning with physics-based hemodynamic correction to non-invasively map 3D vascular geometry to 4D blood-flow fields, preserving physical consistency with reference CFD simulations and clinical stenosis evaluations at rapid inference speed.
Evaluated 10 LLMs and a 4-radiologist reader study across 433 multicenter patients with free-text carotid MRI reports, demonstrating that explicitly encoding modified-AHA staging criteria improves inter-reader agreement and enables effective deployment of a lightweight 8B model.
Grounded in cardiopulmonary hemodynamics — elevated pulmonary vascular resistance and reduced left-atrial venous return — this study establishes 3D pulmonary artery-to-vein volume difference on chest CTA as an interpretable quantitative biomarker for acute pulmonary embolism risk stratification.
Quantifies vertebral artery injury risks associated with ponticulus posticus anatomical variants via 3D pedicle screw trajectory simulation across 141 cervical CTA examinations, establishing a preoperative anatomical screening workflow.
Loading the latest news…
Led by Prof. Yuehua Li together with PIs Ruipeng Zhang, Jianyong Wei, and Zheng Sun, our group brings together radiologists, engineering researchers, postdoctoral fellows, and graduate students inside a high-volume clinical imaging department.
We welcome multicenter collaborations and recruit PhD and master’s students, postdoctoral fellows, and visiting researchers in medical imaging AI, computer vision, and multimodal learning. Motivated undergraduates from SJTU School of Medicine are also encouraged to participate in RBL (Research-Based Learning) and innovation competitions. Please contact [email protected].