MitoEM Dataset: Large-scale 3D Mitochondria Instance Segmentation from EM Images

Donglai Wei¹ Zudi Lin¹ Daniel Franco-Barranco^2,3 Nils Wendt^4* Xingyu Liu^5* Wenjie Yin^1* Xin Huang^6* Aarush Gupta^7*
Won-Dong Jang¹ Xueying Wang¹ Ignacio Arganda-Carreras^2,3,8 Jeff W. Lichtman¹ Hanspeter Pfister¹

¹Harvard University     ²Donostia International Physics Center     ³University of the Basque Country
⁴Technical University of Munich     ⁵Shanghai Jiao Tong University     ⁶Northeastern University
⁷Indian Institute of Technology Roorkee     ⁸Ikerbasque, Basque Foundation for Science

* Works were done as interns at Harvard University

Abstract

Electron microscopy (EM) allows the identification of intracellular organelles such as mitochondria, providing insights for clinical and scientific studies. However, public mitochondria segmentation datasets only contain hundreds of instances with simple shapes. It is unclear if existing methods achieving human-level accuracy on these small datasets are robust in practice. To this end, we introduce the MitoEM dataset, a 3D mitochondria instance segmentation dataset with two 30x30x30 μm³ volumes from human and rat cortices respectively, 3,600x larger than previous benchmarks. With around 40K instances, we find a great diversity of mitochondria in terms of shape and density. For evaluation, we tailor the implementation of the average precision (AP) metric for 3D data with a 45x speedup. On MitoEM, we find existing instance segmentation methods often fail to correctly segment mitochondria with complex shapes or close contacts with other instances. Thus, our MitoEM dataset poses new challenges to the field.

Dataset

(Left) We plot the length versus volume of mitochondria instances for both volumes, where the length of the mitochondria is approximated by the number of voxels in its 3D skeleton. (Right) There is a strong linear correlation between the volume and length mitochondria in both volumes, which is the average thickness of the instance. While the MitoEM-H has more small instances, the MitoEM-R has more large instances with complex morphologies. We sample mitochondria of different length along the regression line and find instances share similar shapes to MOAS in both volumes.

Updates from the Proceeding Version

The MitoEM dataset is 1,986x bigger than the Lucchi dataset, instead of the 3,600x claimed in the paper. We mistakenly doubled the factor by plugging in only the size of the Lucchi test data. Lucchi dataset annotates two sub-volumes (each 165*1024*768 in voxels) from the 5*5*5 um image volume, and each of the two MitoEM volumes has 4096*4096*1000 voxels at the 8*8*30 nm resolution. Thus, the factor should be ((4096*4096*1000)*(8*8*30)) / ((165*1024*768)*(5*5*5)) \approx 1986.
Sec. 2 "Dataset Acquisition": the human data is from the temporal lobe instead of the frontal lobe. The entire human dataset was introduced in [Shapson-Coe et al. 2021]

After the proceeding, we did another round of annotation cleaning and improved the training of our 3D model. Please cite and compare with the following numbers, if you plan to use this dataset. [Descriptions] [Leaderboard Results]

Method	MitoEM-H				MitoEM-R
Method	Small	Med	Large	All	Small	Med	Large	All
U2D-B v2	0.106	0.592	0.563	0.566	0.057	0.450	0.300	0.335
U3D-BC v2	0.426	0.838	0.798	0.804	0.311	0.845	0.803	0.816

Citation

@inproceedings{wei2020mitoem,
  title={MitoEM Dataset: Large-Scale 3D Mitochondria Instance Segmentation from EM Images},
  author={Wei, Donglai and Lin, Zudi and Franco-Barranco, Daniel and Wendt, Nils and Liu, Xingyu and 
  Yin, Wenjie and Huang, Xin and Gupta, Aarush and Jang, Won-Dong and Wang, Xueying and others},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={66--76},
  year={2020},
  organization={Springer}
}

Acknowledgement

This work has been partially supported by NSF award IIS-1835231 and NIH award 5U54CA225088-03.