fidoptIn silico labeling, transmitted light to nucleus fluorescence, 2D
One well of a cell culture in 2D: the middle plane of a transmitted-light z-stack (bright field, phase contrast or DIC) as source, and the maximum projection of the nucleus stain's fluorescence of the same field as target (In silico labeling, Google and the Finkbeiner and Rubin labs).
17 test casescellsource transmitted light, target nucleus fluorescence2D, distances in px
Read before comparing (9)
- Case. The source is the middle plane (z_depth-6) of the well's 13 transmitted-light planes: bright field for the Rubin lab's condition, phase contrast for the Finkbeiner lab's three, DIC for Google's. The target is the maximum projection of the nucleus stain over its own z-stack, confocal or, for the Rubin lab, widefield. 1870 x 2574 to 4686 x 4622 pixels; the files give no pixel size.
- Focus. The middle plane is not each well's focus: the variance of the Laplacian, a sharpness measure, peaks at plane 5 on alicia_2_0_A1 and at plane 9 or later on scott_1_0_r03c04. The provider's own network took all 13 planes.
- Values. Both images are 16-bit as delivered. Every transmitted image is centred near 32,768 (its median 32,100 to 33,150), the provider's normalisation. The targets sit on a background of 9,600 to 15,000, saturate at 65,535 on 0 to 1.2% of the pixels (the confocal conditions), and the Rubin lab's widefield images stay under 50,000. truth.dynamic_range is the 16-bit scale, 65,535.
- Tiles. The Rubin lab's nucleus images are stitched from tiles whose brightness steps at the seams; the seams count as target for every method.
- Alignment. The planes and the stain come from one acquisition on one grid, and nothing is registered; on crops of five test wells the nuclei fall on the cells of the transmitted image.
- Mask. The whole field: a mask made from the target would give the methods the nuclei's places.
- Baseline. On the whole field of the 17 test wells, at a range of 65,535, a constant at the target's median scores an MAE of 3,588 (1,210 to 5,793), a PSNR of 17.8 dB (17.0 to 20.3) and an SSIM of 0.84 (0.66 to 0.97): the dark background it gets right lifts its SSIM, least on the Rubin lab's crowded wells (0.66). The transmitted image itself scores 18,200, 10.4 dB and 0.16.
- Outside numbers. The bucket holds the provider's predictions for every test well (kind PREDICTED), from a network trained on the training wells; the paper reports their correlation with the targets per condition. Any model trained on the training wells has seen the development cases.
- Overlap. No other dataset of the corpus holds these images; LightMyCells_Nucleus_Syn is another deposit of other cells.
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 test set whole: its 17 wells, 3 of the Rubin lab's condition (A in the paper), 4, 8 and 1 of the Finkbeiner lab's three (B, C, D) and 1 of Google's (E). Then 12 development wells, the last of the seeded order of the provider's 119 training wells, one case a well. As in the provider's split, other wells of the same plates are on both sides: 8 of the 12 are of condition C, 2 of A and 2 of B.
- Truth
- target image as released the well's nucleus stain (DAPI) fluorescence as the provider gives it, the maximum projection over its z-stack, 16-bit, on the transmitted-light images' grid, from the same acquisition
- Source
- github.com/google/in-silico-labeling/blob/master/data.md
storage.googleapis.com/in-silico-labeling/paper_data/test_single_channel_images_data_index.html
doi.org/10.1016/j.cell.2018.03.040
github.com/google/in-silico-labeling - Licence
- CC BY 4.0 (the provider's data guide, data.md)
12 cases for tuning, never tested
scott_1_0_r05c07, alicia_2_0_C11, alicia_2_0_E4, kevan_0_7_B8, scott_1_0_r06c06, kevan_0_9_C1, alicia_2_0_H1, alicia_2_0_G11, alicia_2_0_C4, alicia_2_0_B10, alicia_2_0_G10, alicia_2_0_D9