einarolafsson/spacr-example-make-masks

spaCR — Make Masks test data

Ten fluorescence microscopy fields of Toxoplasma gondii parasitophorous vacuoles, each with a curated instance mask. This is the data behind the Load test data… button on the Make Masks screen of spaCR.

Where the images come from

All ten are the lab's own acquisitions. They come from the in-house part (toxoplasma_pv) of spaCR's Toxoplasma vacuole training set, and each one's row in that set's manifest is reproduced below and in manifest.csv. None is a figure from a published paper: the training set's literature part is excluded, and each stem was checked against the literature manifest.

None is a held-out image either. The ten are among the lab-annotated plate wells that the vacuole model was trained on. The eleven wells held out from that training, and every image in a hold-out or curation queue, were excluded.

Because these images are training data, do not use them to score spaCR's Toxoplasma vacuole models. Those models have already seen them.

The ten fields

Stems are plate<N>_<well>_<field>. The name, host and dataset columns are the in-house manifest's row, verbatim. host is ? in that manifest for all ten: the host cell line was not recorded.

Image	name	host	dataset	Size (px)	Vacuoles in the mask	Mask type
plate1_A02_12.tif	plate1_A02_12	?	toxoplasma_pv	1994 × 1994	33	uint16
plate2_E08_13.tif	plate2_E08_13	?	toxoplasma_pv	1994 × 1994	23	uint16
plate2_N04_5.tif	plate2_N04_5	?	toxoplasma_pv	1994 × 1994	109	uint16
plate3_B18_17.tif	plate3_B18_17	?	toxoplasma_pv	2000 × 2000	218	uint16
plate3_I10_4.tif	plate3_I10_4	?	toxoplasma_pv	2000 × 2000	18	int32
plate4_C01_6.tif	plate4_C01_6	?	toxoplasma_pv	2000 × 2000	92	uint16
plate5_G12_7.tif	plate5_G12_7	?	toxoplasma_pv	1994 × 1994	139	uint16
plate6_D16_24.tif	plate6_D16_24	?	toxoplasma_pv	1994 × 1994	45	uint16
plate7_D05_7.tif	plate7_D05_7	?	toxoplasma_pv	1994 × 1994	176	uint16
plate7_N21_6.tif	plate7_N21_6	?	toxoplasma_pv	1994 × 1994	57	uint16

They span seven plates and ten different wells, with 18 to 218 vacuoles per field. Fields with too few objects to be useful, which are mostly camera noise, were left out.

What is here

Path	What it is
<stem>.tif × 10	The images, byte-identical to the training set: single-channel 16-bit TIFF, deflate-compressed, not rescaled.
ground_truth_masks/<stem>.tif × 10	The curated label masks, byte-identical to the training set. Each vacuole is a distinct integer and 0 is background. They are stored as the training set stores them, so most are uint16 and one is int32.
manifest.csv	The ten manifest rows, plus each field's size, pixel types, object count and the SHA-256 of both files.
spacr-example-make-masks.tar	All of the above in one uncompressed archive, plus an empty masks/ folder. This is what spaCR downloads.

The masks are ground truth

The masks in ground_truth_masks/ are the curated labels the vacuole model was trained on. Make Masks does not read that folder: it looks for drafts in masks/. The images therefore open unsegmented, so there is something to segment and edit, and the truth sits beside them for comparison.

Using it in spaCR

In Make Masks, press Load test data…. The first press downloads spacr-example-make-masks.tar (about 38 MB) into ~/.cache/spacr/example_data/make_masks_toxo_pv/, unpacks it, and opens the folder on its first image by name, plate1_A02_12.tif. Later presses open the cached copy without downloading anything. Masks you save go to masks/ in that folder and never touch ground_truth_masks/.

To use the data outside the button, download the archive, unpack it, and open the folder with Open folder… in Make Masks, or run spacr-make-masks --folder <that folder>.

License

MIT, the same as spaCR's other example datasets.