EN /

Projects

From clinical AI agents to single-cell foundation models — selected work across biomedical computing, multimodal learning, and applied ML.

Featured
···
All Projects

RST2G: Breast Cancer Segmentation

Spatiotemporal transformer with graph fusion for DCE-MRI breast tumor segmentation. Collaboration with Tsinghua SIGS (Prof. Peiwu Qin's team). Achieves state-of-the-art on 2 public breast DCE-MRI datasets, outperforming existing 2D, 3D, and 4D approaches.

Published in Cyborg and Bionic Systems · 3rd Author

Medical ImagingTransformerGraph Neural Network
Read Paper (PDF) → DOI ↗

ImmuneWorld: Foundation-Model World Simulator for TIME

First foundation model-based world simulator for the tumor immune microenvironment. A 12-layer trajectory-aware Transformer autoregressively pre-trained on 12.4M single cells, achieving Pearson r=0.914 on cell-state transition prediction. Cross-cancer transfer module lifts ICB response AUC to 0.891 across 14 cohorts and 7 cancer types.

Under Review at Nature Communications · 4th Author · Collaboration with BIT

Foundation ModelWorld SimulatorSingle-CellImmunotherapy

Smart Power Load Forecasting

Regional power demand prediction system using ensemble ML (LSTM, LightGBM, Prophet) and LLM semantic analysis. Built multi-source data fusion models achieving MAPE < 3%, R² > 0.97 across multiple scenarios.

Huadian Energy Research Center · Project Manager

Time SeriesLSTMLightGBMEnergy

IMU Pose Estimation with Diffusion Models

Initialization-free human pose estimation from IMU sensors using Diffusion Models with physics-based dynamics constraints. Designed for rehabilitation medicine and sports monitoring applications.

HKUST(GZ) SIIE Lab · Research Assistant

Diffusion ModelPose EstimationIMU

CA Cancer Journal · Global Bone Tumor Epidemiology Review

A multi-center evidence synthesis targeting CA: A Cancer Journal for Clinicians (IF ~500+, the highest-impact medical journal in the world). Integrates bone-sarcoma imaging, pathology, and follow-up data from 6 Class-A hospitals to build a cross-region, cross-age, cross-subtype global epidemiological portrait and clinical decision pathway.

In Preparation · Co-First Author (later in order) · With Dr. Mingyang Jiang (NUS) · Target submission Summer 2026

EpidemiologyMulti-CenterBone SarcomaClinical Review

Multi-Center Bone Sarcoma Imaging Platform

A bone-sarcoma imaging / pathology / follow-up data platform jointly built by 6 Class-A hospitals. Responsible for compliant ingestion, cross-center schema alignment, and a tri-modal (imaging-pathology-clinical) data pipeline that serves as the data backbone for multi-center evidence synthesis.

Infrastructure work · 6 Class-A hospitals · Powers the CA Cancer Journal review and downstream modeling

Data EngineeringMultimodalImagingClinical Data

Shenzhen Xinyu Zhilian Tech Co., Ltd.

A tech company I co-founded, working in the ed-tech space. Closed an end-to-end loop from product definition and engineering to operations and go-to-market — independently. Cumulative net profit of 1M+ RMB to date. A real-world experience that runs both the "research" and "product" mental models on the same brain.

Co-founder · 2023 – Present · Cumulative net profit 1M+ RMB

EntrepreneurshipEd-TechProductCommercialization