I am a final-year PhD candidate in Physics (Medical AI) at PSI & ETH Zurich, specializing in generative models and physics-informed deep learning for medical imaging and treatment planning.
I develop and evaluate deep generative and representation-learning models for high-dimensional imaging and spatiotemporal data. Using Diffusion Models, Flow Matching, Implicit Neural Representations, and Transformer architectures, I build end-to-end ML pipelines for tasks such as high-fidelity image synthesis, real-time motion tracking, and adaptive dose prediction in clinical workflows.
I am embedded within the Center for Proton Therapy (CPT). My doctoral research is supervised by Prof. Antony Lomax and Dr. Ye Zhang, and is supported by the SNSF project EPIC-4DAPT (Grant No. 212855).
Prior to Switzerland, I received my M.S. (2023) and B.S. (2020) degrees from Tsinghua University. During my master's in the Department of Automation, I focused on 3D generative vision and multimodal video understanding under the guidance of Prof. Jiwen Lu and Prof. Jianjiang Feng.
A proof-of-concept neural engine that predicts individual proton pencil beam dose distributions directly from MRI. A Transformer-based model operating on beam’s-eye-view patches achieves Monte Carlo-comparable accuracy (median 99.8% gamma pass rate at 1 mm/1%) while reducing computation time to about 3 ms per beam (a 600x speedup over FRED MC on GPU).
A DSBM framework for high-fidelity MR-to-CT synthesis that models an entropic optimal transport (Schrödinger bridge) between MR and CT distributions. Unlike standard diffusion models starting from pure noise, DSBM leverages MRI as a structural prior to preserve geometric fidelity and dosimetric accuracy.
Adapts Instant Neural Graphics Primitives (NGP) and multi-resolution hash encoding to deformable medical image registration. Enables on-the-fly motion extraction with sub-second inference for 4D dose accumulation, addressing motion management bottlenecks.
A continuous function representation f(x, y, z, t) for 4D-MRI that eliminates binning artifacts and enables high-quality volumetric MRI reconstruction at arbitrary respiratory phases, critical for tracking irregular breathing motion.
* indicates equal contribution.
Presents the MIND neural dose engine that predicts individual proton pencil beam dose distributions directly from MR images, achieving a median 99.8% gamma pass rate at 1 mm/1% while reducing computation to about 3 ms per beam.
Develops an xLSTM-based CBCT neural dose engine (CBCT-NN) that predicts proton dose directly on CBCT images, achieving 95.1 ± 2.7% gamma pass rates at 2 mm/2% and under-3-minute computation for complete head-and-neck plans, enabling adaptive workflows without synthetic CT or complex correction pipelines.
Introduces DSBM, modeling an entropic optimal transport (Schrödinger bridge) between MR and CT distributions to achieve superior geometric fidelity and dosimetric accuracy in MR-only proton therapy.
Adapts Instant Neural Graphics Primitives and multi-resolution hash encoding to enable on-the-fly motion extraction with sub-second inference for accurate 4D dose accumulation.
Proposes a continuous spatio-temporal representation f(x, y, z, t) for 4D-MRI, enabling reconstruction at arbitrary respiratory phases and reducing binning artifacts.
Proposes Diffusion-SDF, combining an SDF autoencoder with voxelized diffusion for text-to-shape synthesis, and extends it to text-conditioned shape completion and manipulation.
Introduces Bridge-Prompt, a prompt-based framework for ordinal action understanding that models semantics across multiple adjacent correlated actions in instructional videos.
Proposes an uncertainty-aware representation learning framework to capture transitional expressions between action periods, improving robustness in action segmentation.
Presents an Order-Constrained Representation Learning (OCRL) approach with a StepNCE contrastive loss to predict future actions from partially observed instructional videos in a weakly-supervised manner.