Pak Hei (Hugo) YEUNG 楊栢熙

I am a Presidential Postdoctoral Fellow at the Nanyang Technological University (NTU), Singapore, working with Prof. Jagath Rajapakse.

I am also an associate researcher in the Oxford Machine Learning in NeuroImaging Lab at the University of Oxford, where I obtained my DPhil (i.e. PhD) under the supervision of Prof. Ana Namburete and Prof. Weidi Xie, generously supported by The R C Lee Centenary Scholarship.

I received my BEng in Medical Engineering from the University of Hong Kong, where I was supervised by Dr. Wei‐Ning Lee for my final year project.

Email  /  Google Scholar  /  Twitter  /  Github  / 

I'm interested in medical image analysis, machine learning, ultrasound imaging, and computer vision. Please check my Google Scholar for the most updated and complete list of publications.

* stands for co-first or co-last authorship.

A model-agnostic approach for transferring knowledge from 2D natural images pretrained models for 3D medical image segmentation with only a few labeled data.

An adaptive federated prompt tuning framework to enhance CLIP's generalization in multi-domain scenarios by disentangling semantic and domain-specific features with dual-prompt design.

An automatic ensemble framework to generate subcortical segmentation labels for CT scans by leveraging existing MRI-based models, providing the first publicly available annotated CT dataset in the field.

A novel test time adaptation method for improving 3D subcortical segmentation of fetal brain in ultrasound images.

A neural representation framework using tensor-rank decomposition to enable rapid 3D volume reconstruction from 2D freehand scans.

An uncertainty-aware deep learning model for automated 3D plane localization in 2D fetal brain images.

Reconstructing a 3D volume from 2D freehand ultrasound images by training a deep network to implicitly represent the volume.

Sensorless 3D localization of 2D freehand scans. Trained with just 3D volumes, the model can be adapted to any domains (e.g. machines) in an unsupervised manner.

Reconstructing a 3D volume from 2D freehand ultrasound images by training a deep network to implicitly represent the volume.

Trained with just raw 3D volumes, a single Sli2Vol model can be used to propagate a single-slice annotation to the whole 3D volume, for any structures across different modalities.

A network, trained with just registered volumes, to predict the 3D location of 2D freehand ultrasound fetal brain images and video.