All publications

Conference paper / 2022

Eigenvector Grouping for Point Cloud Vessel Labeling

Patryk Rygiel, Maciej Zięba, Tomasz Konopczyński

Geometric Deep Learning in Medical Image Analysis (GeoMedIA) · Published

A U-Net coronary segmentation is converted to a point cloud, labeled with Eigenvector Grouping and refined with component voting.
The workflow refines an initial coronary segmentation using PointNet++ with Eigenvector Grouping and component-wise voting. Figure 1 · Rygiel et al. (2022) · © 2022 P. Rygiel, M. Zięba and T. Konopczyński

Neighborhoods shaped for vessels

Point-cloud networks often aggregate information from generic local neighborhoods. Eigenvector Grouping tailors this operation to the elongated structure of blood vessels. A component-wise voting step then encourages consistent labels within connected vessel segments.

The method is evaluated on coronary CT angiography, including cases with interrupted segmentations. The official implementation accompanies the PMLR paper.

Presented as a poster at the GeoMedIA workshop in Amsterdam in November 2022.

Cite this work

@inproceedings{eigenvectorgrouping2022,
  title = {Eigenvector Grouping for Point Cloud Vessel Labeling},
  author = { Patryk Rygiel and Maciej Zięba and Tomasz Konopczyński },
  year = {2022},
  booktitle = {Geometric Deep Learning in Medical Image Analysis (GeoMedIA)},
  volume = {194},
  pages = {72--84},
  publisher = {PMLR},
  url = { https://proceedings.mlr.press/v194/rygiel22a.html }
}