EchoPilot
One click. One category. No training.
Stable ultrasound video segmentation with frozen foundation models.
Ph.D. student
Ruiqiang Xiao （肖睿强）
Medical image analysisVideo understandingMultimodal AI
I’m a second-year Ph.D. student at the Hong Kong University of Science and Technology (Guangzhou), advised by Prof. Lei Zhu and Prof. Kaishun Wu. I’m currently an intern at JD.com.
I received my master’s degree from the Hong Kong University of Science and Technology (HKUST) and my bachelor’s degree from Southeast University (SEU).
01 Research in focus
Selected research
Anatomy in motion
Real-time anatomical recognition in liver surgery.
Video segmentation evaluated across centers and with surgeons.
02 Recent updates
News
Research, publications
and milestones.
Our work on real-time surgical anatomy recognition is published in npj Digital Medicine.
Cardio-mmFlow, our physics-informed ECG-to-radar synthesis framework, appears at ICML 2026.
We share HyperCBM, a label-efficient framework for interpretable medical image diagnosis.
Earlier updates
IRFusionFormer receives the Best Paper Award at the BMVC DIFA Workshop.
03 Research record
Publications
20266 papers
EchoPilot: Training-Free Ultrasound Video Segmentation via Scale-Space Semantic Prompting and Reliability-Gated Memory
One point and an anatomical category name guide frozen foundation models, with scale-aware prompting and gated memory for stable ultrasound video segmentation.
Discriminative Synergistic Temporal Diffusion for Ultrasound Video Thyroid Nodule Segmentation
Timestep-dependent motion and texture guidance help a diffusion model delineate small thyroid nodules and ambiguous boundaries in ultrasound videos.
Real-time anatomy recognition in laparoscopic liver resection using video segmentation AI model
A Mamba-based video segmentation model recognizes surgical anatomy in real time, evaluated on multicenter liver surgery videos and with surgeons.
Cardio-mmFlow: A Gaussian-Prior-Free Physics-Informed Flow Matching Framework for Electrocardiogram to mmWave Radar Synthesis
Physics-informed flow matching translates ECG into personalized mmWave radar signals to support contactless cardiac monitoring with synthetic training data.
Learning Label-Efficient Interpretable Medical Image Diagnosis via Semi-supervised Hypergraph Concept Bottleneck Model
Dual-level hypergraphs model relationships between clinical concepts and generate pseudo-labels for interpretable diagnosis with fewer expert concept annotations.
OsmT: Bridging OpenStreetMap Queries and Natural Language with Open-source Tag-aware Language Models
Tag-aware language models generate and explain OpenStreetMap queries, using retrieval augmentation to connect natural language with structured geospatial data.
20251 paper
HRVVS: A High-resolution Video Vasculature Segmentation Network via Hierarchical Autoregressive Residual Priors
Autoregressive visual priors and a dynamic memory decoder preserve vessel detail across high-resolution liver surgery video frames.
20242 papers
MatchSeg: Towards Better Segmentation via Reference Image Matching
CLIP selects relevant reference images, while joint attention transfers support-set information for medical segmentation and cross-domain generalization.
IRFusionFormer: Enhancing Pavement Crack Segmentation with RGB-T Fusion and Topological-Based Loss
RGB and thermal features are fused with cross-modal supervision and a topology-aware loss to preserve crack connectivity under challenging imaging conditions.
04 Background
Education
& experience
CurrentEducation
HKUST (Guangzhou)
Second-year Ph.D. student
Advised by Prof. Lei Zhu and Prof. Kaishun Wu
CurrentIndustry
JD.com
Intern
Let’s connect