Ortho-plane inference is used for isotropic voxel EM data. Independent stacks of 2D predictions on XY, XZ, and YZ planes are averaged into a final 3D segmentation. This can introduce artifacts like "cross-hatching" that typically would require manual cleanup. Training a separate DL model with weak supervision (bootstrapping) on the ortho-plane inference output can significantly improve 3D segmentation quality.
-
boostrap2.5d: Source code to run orthoplane inference and weakly supervised training.
-
APEER module: An APEER module to easily apply this technique on any 3D binary segmentation task.
-
Colab notebook: Google Colab notebook that demonstrates an application for mitochondria and lysosomes.
If you find any of these resources useful in your work, please cite:
@article{conrad_lee_narayan_2020,
title={Enforcing Prediction Consistency Across Orthogonal Planes Significantly Improves Segmentation of FIB-SEM Image Volumes by 2D Neural Networks.},
volume={26},
DOI={10.1017/S143192762002053X},
number={S2},
journal={Microscopy and Microanalysis},
publisher={Cambridge University Press},
author={Conrad, Ryan and Lee, Hanbin and Narayan, Kedar},
year={2020},
pages={2128–2130}
}