Try it on public data¶
The Image Data Resource (IDR) publishes reference
image sets from published studies. Many studies are also available as OME-Zarr on
public S3 storage, which OMERO Annotate AI reads directly (source: zarr). No OMERO
login is needed, and because nothing can be written to the IDR, the training set is
kept on your computer (storage: local).
The example used below is HeLa nuclei (histone H2B-GFP) from the MitoCheck screen,
study idr0013: 2D
time-lapses of 1024 × 1344 pixels, with nuclei about 25 pixels across, a good size for
micro-SAM's vit_b_lm model.
With the demo notebook¶
The IDR demo notebook (marimo) checks a screen for OME-Zarr plates, creates the training set, opens napari and trains a short micro-SAM run, with everything pre-filled.
With a YAML file¶
The same example as a config,
idr0013-mitocheck-zarr.yaml:
name: idr0013-mitocheck-nuclei
source: zarr
zarr:
uris:
- https://uk1s3.embassy.ebi.ac.uk/idr/zarr/v0.4/idr0013A/3451.zarr
level: 0 # full resolution
fields: first
wells: [A1, B2, C3, D4]
spatial_coverage:
channels: [0] # H2B-GFP: nuclei
timepoint_mode: random
n_timepoints: 1 # one random frame of each time-lapse
training:
train_n: 2
validate_n: 1
seed: 42
storage: local
Load it in the annotate-from-YAML notebook, or in Python:
from omero_annotate_ai import TrainingSet, load_config_from_yaml
ts = TrainingSet.create(load_config_from_yaml("idr0013-mitocheck-zarr.yaml"))
ts.annotate(note="first try on IDR data")
Only metadata is read when the set is created; the pixels of each chosen plane are downloaded when it is annotated.
Finding other data¶
Only images converted to OME-Zarr can be used, and often only some plates of a study are converted (in idr0013A, for example, plate 3451 but not its neighbours). The IDR OME-NGFF samples page lists what is available (studies with plates include idr0001A, idr0013A, idr0056B, idr0072B and idr0128E). A plate's OME-Zarr address is
where the study folder is the study name with its screen letter (idr0013A) and the
plate ID is the one the IDR shows. Plate IDs of a screen come from the IDR web API, which
needs no login (adapted from the IDR's
IDR API example notebook):
import requests
screen_id = 1101 # idr0013-neumann-mitocheck/screenA
plates = requests.get(
"https://idr.openmicroscopy.org/webclient/api/plates/",
params={"id": screen_id},
timeout=30,
).json()["plates"]
for plate in plates[:5]:
url = f"https://uk1s3.embassy.ebi.ac.uk/idr/zarr/v0.4/idr0013A/{plate['id']}.zarr"
ok = requests.head(url + "/.zattrs", timeout=15).status_code == 200
print(plate["id"], plate["name"], url if ok else "no OME-Zarr")
Pick the channel from the study's description (the OME-Zarr metadata lists channel
labels too), choose wells with zarr.wells or a metadata CSV with
well filters, and use zarr.level to work at a lower
resolution when objects are large.
What you get¶
The set's folder (~/omero_annotate_ai/local/<id>_<name>/) holds:
annotation_input/ the images you annotated, as TIFF
annotation_output/ your masks, <unit id>_mask.tif
tracking_table.csv which units were done, and their split
config.yaml the set's config, sessions and notes
These go straight into training.