fidopt Interactive view
Fluorescence microscopy, low to high exposure, three specimens, synthesis

CARE Planaria, low-exposure to high-exposure fluorescence, 3D

A spinning-disk confocal stack of planarian flatworms stained for their nuclei, taken at 58 times less light (source) and at full light (target), the two conditions interleaved plane by plane in one acquisition (CARE, Weigert et al. 2018).

20 test casesflatwormsource low exposure, target high exposure
Look at the images
Read before comparing (8)
  • Pairs. On a spinning-disk confocal with a 30x silicon-oil objective, the provider imaged planarians stained for their nuclei (RedDot1) in four conditions interleaved plane by plane: GT at 2.31 mW and 30 ms, and C1 to C3 at less light. The source is C2, 0.12 mW and 10 ms, 58 times less light than GT, the paper's 60-fold example; C1 and C3 are not used.
  • Stacks. The 20 test stacks are 1024 x 1024 pixels of 26 to 95 planes. Sixteen (EXP278 ... m0001 to m0016) are tiles of one fixed worm, six rows of two or three, overlapping their neighbours by about 158 pixels (found by phase correlation of their projections), so a fifth of their voxels repeat another tile's; the other four are two stacks each of two live animals (EXP280 N2 and N3), the two stacks of a worm not overlapping. The cases are not twenty independent samples.
  • Values. Both stacks are uint16 camera counts as delivered, on an offset near 490. The low-exposure stack peaks near 660 to 950 at its 99.9th percentile and is mostly noise; the target's 99.9th percentile runs from 6,000 to 30,400, and truth.dynamic_range, 18,225, is that percentile over every voxel of the 20 stacks. The files give no voxel size: the unit is the voxel.
  • Alignment. The provider interleaved the conditions plane by plane to keep the stacks registered; nothing is registered here.
  • Mask. The whole stack, as the paper scores it: much of the tiles is empty slide.
  • Baseline. On the whole stack of the 20 test cases, at a range of 18,225, a constant at the target's median scores an MAE of 656 (99 to 1,446), a PSNR of 23.1 dB (16.6 to 33.3) and an SSIM of 0.70 (0.38 to 0.96); the low-exposure stack itself scores 999 (111 to 3,029), 21.9 dB and 0.63, worse than the constant on all 20, being on another scale.
  • Outside numbers. The paper scores its network, NLM and other denoisers on these stacks with NRMSE and SSIM after a normalisation of its own, not on this scale; its training patches come from other stacks of the same experiments.
  • Overlap. No other dataset of the corpus holds images of this deposit.

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

Case by case

Click a score to look at that result: the target image, and over it the method's synthetic image. Underlined scores have their image; for the others you see the case without the result.

About the data

Test cases
The provider's 20 test stacks whole; its training data are patches cut from other stacks, so there is no development case. The stacks are not 20 subjects: 16 are overlapping tiles of one fixed worm and 4 are two stacks each of two live worms.
Truth
target image as released the stack at high laser power and exposure (GT, 2.31 mW and 30 ms), interleaved plane by plane with the low-exposure one in one acquisition on one microscope, on the same grid, not registered
Licence
CC0 1.0 (the Edmond deposit's licence field)