王栩成
I am a master’s student in Biomedical Informatics at Harvard Medical School, advised by Pranav Rajpurkar. My research focuses on developing and evaluating clinical AI agents and embodied AI for medicine—systems that perceive, reason, and act alongside clinicians, from simulated hospitals to surgical robots.
My work has been accepted at conferences including CHIL, MLHC, and ML4H, and in journals including Science and the American Journal of Transplantation.
I earned a BS in Computer Science from Washington University in St. Louis, with a second major in Financial Engineering and a minor in Statistics. There, I worked with Jin Zhang on AI for radiation oncology and Ulugbek Kamilov on diffusion models for medical image synthesis and reconstruction.
Outside the lab, I enjoy playing and listening to music , road trips and off-roading , and skiing .
Our papers SimRetract and ECGDreamSuite were accepted to the ML4H 2026 Proceedings track, and SurgeCount to the Findings track.
Our paper, Abandonment of Safe Medical Advice by Large Language Models Under Patient Resistance, has been accepted for publication in Science. The paper will be available upon publication.
Received First Place, Best Poster Presentation at DBMI Science Day, Department of Biomedical Informatics, Harvard Medical School.
Our paper, Donor Care Unit Availability and Organ Procurement Organization Performance in Lung Transplantation, has been accepted and is now online in the American Journal of Transplantation.
Presented our paper, Imitation Learning for Robot Assistance in Open Surgery: A Multi-Policy Evaluation on Suture Following, at the Machine Learning for Healthcare Conference (MLHC) in Baltimore.
Our paper, Imitation Learning for Robot Assistance in Open Surgery: A Multi-Policy Evaluation on Suture Following, was accepted to the Machine Learning for Healthcare Conference (MLHC 2026).
Presented ReXSonoVQA at the Conference on Health, Inference, and Learning (CHIL) in Seattle.
Our paper, ReXSonoVQA: A Video QA Benchmark for Procedure-Centric Ultrasound Understanding, was accepted to the Conference on Health, Inference, and Learning (CHIL 2026).
Joined Dr. Pranav Rajpurkar’s lab at Harvard Medical School.
Presented our work on bladder-aware PET segmentation and eRT-RAG for radiation oncology at the Radiation Research Society (RRS) Annual Meeting.
Joined Harvard Medical School as a master’s student in Biomedical Informatics.
Learns retractor take-over entirely from simulated task demonstrations and transfers zero-shot to the real robot while preserving the surgeon-established retraction.
More info will be available.
An LLM-driven clinical world model for interactive simulation and agent evaluation in emergency-department settings, with a focus on patient prioritization, demographic equity, and care management under resource constraints.
Evaluates multiple imitation-learning policies for coordinated surgeon–robot manipulation and tests the resulting system in a live running-suture trial.
Tests whether video-language models can understand ultrasound procedures, including action goals, troubleshooting, and causal planning across six clinical domains.
Frames multimodal reconstruction and synthesis as one inverse problem and develops inference-time diffusion samplers for flexible MRI, CT, and PET recovery.
Robot learning · Human–robot collaboration
Developing a portable multi-arm assistive system for open surgery, spanning cardiac procedures. The work connects learned manipulation skills, safe handover, and VLM-supervised execution to support surgical teams.
Hierarchical control · Multimodal perception · 3D reconstruction
Developing a hierarchical vision-language-action framework with a VLM planner and an imitation-learning 6-DOF controller. Synchronized ultrasound, RGB-D, robot kinematics, and expert narration support learning how to scan; probe-pose-conditioned 3D organ reconstruction helps interpret the scanning procedure.
Deformable registration · Synthetic training data · Anatomical supervision
Developed a semi-supervised, segmentation-aware motion-correction method for in-vivo two-photon calcium imaging. Building on VoxelMorph and SynthMorph, the approach uses synthetic neuron images and auxiliary segmentation supervision to align frames while preserving neuron shapes and locations, supporting downstream neural activity analysis with limited real training data.
3D vision · Cross-view reconstruction
Synthesizing street-view panoramas from overhead imagery and nearby views. Height-derived structure maps and NeRF-style positional encoding guide geospatial cross-attention, combining scene geometry with appearance features to condition ControlNet and Stable Diffusion.
, and skiing 
.