fidopt Interactive view
Electron microscopy, organelles, segmentation

DeePiCt S. pombe cryo-electron tomograms

A cryo-electron tomogram of a FIB-milled lamella of a fission yeast cell (S. pombe), at 2.7 nm a voxel, with its organelles and cytoplasm segmented or corrected by hand (DeePiCt, CryoET Data Portal).

10 test casescellimage EM
Look at the images
Read before comparing (10)
  • Truth. Masks drawn by hand in Amira on the phase-plate tomograms (vpp_), and on the defocus tomograms (defocus_) a 2D network's prediction, trained on those, corrected by hand. The paper does not say who drew, whether a second person checked, or how well two people agree. On the defocus tomograms a method close to DeePiCt's 2D network may be favoured where the correction left its outline.
  • Images. Tilt series over at most -50 to +58 degrees, as little as -30 to +30 on a defocus tomogram, reconstructed by weighted back projection, neither filtered nor denoised. The missing wedge smears every structure along z, and a membrane seen edge-on from above fades. The intensities are the reconstruction's; the cell's matter is dark.
  • Level. A case is the portal's level 1, half the resolution: voxels of 2.696 nm, 480 x 464 in the plane. The full resolution holds several times the 2^26 voxels a segmentation case of the corpus can. Along z a case keeps the annotated slab, 74 to 152 slices on the test cases; the lamella is tilted in it, with no cell above or below.
  • Labels. One label a voxel: where the masks overlap, 0.4 to 1.3% of a tomogram's labelled voxels, the later label of the list wins, the cytoplasm losing to every organelle. Organelle masks are filled; the cytoplasm is the cell's matter outside them, ribosomes included. The nuclear envelope keeps its own label, where OpenOrganelleEM counts it with the nucleus.
  • Left out. The membrane mask, which runs through the organelles; the ribosome and fatty acid synthase positions; vpp_TS_0008's one 'membrane-enclosed lumen' mask, background here; the portal's later depositions 10301, 10303 and 10351. Two one-voxel Golgi masks (vpp_TS_0004, vpp_TS_0005) are not read, and defocus_TS_034's cytoplasm mask keeps its artefact on each slice's first row and column.
  • Absent structures. A mask the deposition does not give is taken as a structure not there; nothing says whether a structure seen was left undrawn. Over the ten test tomograms the cytoplasm and vesicles are in all ten, down to the multivesicular bodies and lipid droplets in two. The mean is over the structures each case holds, so rare ones weigh on two or three cases.
  • Tolerance. Distances are in nanometres and the surface tolerance is 1 nm, under half a voxel: the surface Dice asks for the truth's own outline. The truth with each structure grown by one voxel scores a Dice of 0.92 and a surface Dice of 0.34 over the test tomograms.
  • Exposure. All of it is public with its truth: DeePiCt's networks were trained on these tomograms, and a network trained on the portal's depositions 10000 or 10351 has seen the test cases. Several tomograms may come from one lamella or one cell; the deposition does not say.
  • Overlap. No other dataset of the corpus holds images of the CryoET Data Portal or of EMPIAR-10988.
  • Outside numbers. None: the paper scores organelles against cytosol by a precision-recall curve on sampled voxels, not by structure.

No method has a result on every case of this dataset yet.

Case by case

Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.

About the data

The first test case of DeePiCt S. pombe cryo-electron tomograms
The first test caseClick to look at every case.
Test cases
10 test tomograms drawn with seed 2026 among 19, five acquired with defocus only and five with a Volta phase plate; the 9 others are the development cases. The twentieth, TS_0010 of the phase-plate set, is not a case: its masks are 928 voxels wide for a tomogram of 960, on no grid of the tomogram.
Truth
contours by experts the deposition's organelle masks, one a structure and a tomogram, drawn by hand in Amira on the ten phase-plate tomograms and predicted by a 2D network trained on those, then corrected by hand, on the ten defocus tomograms; the corpus merges them into one map at the portal's half resolution, and leaves the membrane mask out
Licence
CC0 (the CryoET Data Portal's data submission policy: data published on the portal is open to anyone 'subject to a CC0 license'; de Teresa-Trueba et al., Nature Methods 2023, is the reference to cite)
9 cases for tuning, never tested

defocus_TS_026, defocus_TS_029, vpp_TS_0007, defocus_TS_043, vpp_TS_0008, defocus_TS_034, vpp_TS_0002, defocus_TS_027, vpp_TS_0004