CV

Education

Sichuan University, PhD in Artificial Intelligence (Direct PhD Program), Sep 2022 - Present

• College of Computer Science, Artificial Intelligence Applications
• Advisors: Prof. Zhang Yi (IEEE Fellow)

Sichuan University, B.Eng. in Computer Science and Technology, Sep 2018 - Jun 2022

• College of Computer Science

Research Projects

ChestWorld: Chest CT World Foundation Model Project Lead | Intelligent Medical Center, Sichuan University Jun 2026 - Present

• Responsibilities and data infrastructure: Lead pretraining strategy, patient-level data isolation, and downstream interfaces for 118,591 chest CT scans; standardize DICOM parsing, HU calibration, lung/mediastinal windowing, resampling, and abnormal-series filtering with versioned patient/study/series indices.
• Multiscale world modeling: Combine dual-window, multiple-physical-scale inputs with scale-adaptive tokenization, JEPA latent prediction, and Continuous Thought Machine (CTM) disentanglement of task-invariant and task-specific representations; connect downstream heads through adapters and a task router.
• Region-level self-supervision: Extract features with frozen SAM 2 Hiera-L and use similarity graphs and union-find to generate up to 8 class-agnostic 3D connected regions per scan; mask region unions and predict EMA target representations to reduce reliance on manual lesion boxes.
• Training objectives and extensions: Organize P0 dense visual prediction, P1 DINO global preservation, P2 cross-window/cross-scale consistency, and P3 lesion awareness with hard negatives; reserve P4/P5 interfaces for future longitudinal CT change prediction and image–report concept alignment.
• Training and evaluation: Configure rolling checkpoints, representation-collapse monitoring, and frozen probe / LP-FT on 4×RTX 4090 GPUs; use patient-level isolation, 3 random seeds, bootstrap confidence intervals, and sensitivity constraints for evaluation.

Multimodal Radiology Report Generation Agent Algorithm Lead | Intelligent Medical Center, Sichuan University Sep 2025 - May 2026

• ReAct workflow: Designed an observe–plan–call–verify loop for CT / MRI / X-ray, composing modality routing, image quality control, evidence grounding, report drafting, and fact checking as expert nodes.
• Evidence traceability: Connected HeatRAD suspicious-slice discovery with Vol-Nav organ anchoring and PPO lesion navigation for CT; standardized entities, locations, confidence, and provenance across modalities in an evidence ledger to constrain LLaMA3-8B generation.
• Tool validation and human review: Built a typed tool registry with schema validation, timeouts, retries, fallback, and human review; checked entity coverage, negation, laterality, and anatomical consistency, and recorded replayable execution traces with human checkpoints for low confidence, tool conflicts, and final sign-off.
• Offline regression: Maintained fixed case sets by modality and abnormality type to assess entity coverage, evidence hits, negation/laterality consistency, tool success, and format compliance; compared execution traces for model, prompt, and tool upgrades.
• Data and deployment: Curated 28,752 CT / MRI / X-ray image–report pairs with hospital collaborators through entity-level correction, cross-physician consistency review, and rule-based anomaly detection; trained with LoRA rank 64 + 8-bit on 4×RTX 4090 GPUs and packaged a PACS-compatible inference service.

Patient-Level Risk Stratification System for Lung Metastases Algorithm Lead | Intelligent Medical Center, Sichuan University Jun 2024 - Jun 2025

• Complete algorithm pipeline: Built DICOM quality control and resampling, ForeMatch semi-supervised 3D segmentation, IBSI-standard radiomics and Graph Mamba deep representations, and patient-level risk calibration into a complete pipeline.
• Robust modeling: Standardized feature definitions with PyRadiomics and addressed center/device batch effects with ComBat; aggregated multiple lesions per patient, used nested patient-level splits and threshold calibration to avoid leakage, and organized interpretable risk evidence with SHAP.
• Layered quality control: Checked input-series completeness, slice thickness/spacing, and resampling parameters; segmentation volume, connected components, and boundary anomalies; and feature NaNs, constants, and distribution drift, routing critical failures to human review.
• Services and version management: Separated ingestion and de-identification, segmentation, feature extraction, risk inference, and result delivery into components linked by case task IDs; bound model weights to preprocessing and feature versions for reruns, staged replacement, and cross-center troubleshooting.
• DICOM/PACS integration design: Recorded an audit trail from source series to risk estimates and designed MONAI Deploy interfaces for DICOM Study/Series ingestion, DICOM SEG/SR output, PACS delivery, and low-confidence human review.

Skills

• Programming: Python, PyTorch, C/C++, LaTeX
• Research: Medical Image Analysis, Deep Learning, Computer Vision
• Languages: Chinese (Native), English (Proficient)

Awards & Honors

• First Prize, Sichuan Province "Challenge Cup" Competition (2026)
• Sichuan Province Young Doer (2025)
• Bronze Medal, National Postdoctoral Innovation and Entrepreneurship Competition (2023)
• First-Class Doctoral Fellowship, Sichuan University (2022)