Baiding Chen
← All projects

CryoPASS: Domain Adaptation for Cryo-ET Segmentation

Oct 2025 – Jul 2026 · Research Assistant (first author) · Carnegie Mellon University · Prof. Min Xu

Segmenting proteins in real cryo-electron tomograms without any real voxel labels, by adapting models trained on simulated data while avoiding the collapse that sparse foregrounds usually cause.

Problem

Simulated cryo-ET data comes with perfect labels, but real tomograms look different and proteins occupy a tiny fraction of the volume. Standard unsupervised adaptation tends to collapse and predict background everywhere.

Approach

CryoPASS keeps the foreground alive with a foreground-balance loss, narrows the appearance gap with Fourier-domain style alignment, and refines training with prototype-based pseudo-labels.

Results

The method improves on the strongest baseline by 1.8× and recovers 78% of the fully supervised upper bound, while being the most stable method across random seeds. The paper is under review at AAAI.

Related publications

  • Under review
    CryoPASS: Collapse-Free Unsupervised Domain Adaptation for Sparse-Foreground Cryo-ET Segmentation
    Baiding Chen, Xiaolong Wu, Xi Xiao, Mian Wu, Qiayi Zha, Haoran Li, Johan Barthelemy, Tianyang Wang, Xingjian Li, Min Xu
    Under review at AAAI, 2026
    Details
    Main figure of “CryoPASS: Collapse-Free Unsupervised Domain Adaptation for Sparse-Foreground Cryo-ET Segmentation”
    CryoPASS: Collapse-Free Unsupervised Domain Adaptation for Sparse-Foreground Cryo-ET Segmentation