Datasets:
File size: 3,737 Bytes
51b3888 117a549 51b3888 e812312 51b3888 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | ---
license: mit
task_categories:
- image-classification
tags:
- spacr
- microscopy
- annotation
- toxoplasma
pretty_name: spaCR Annotate and Classify example data
---
# spaCR — Annotate and Classify example data
Example input for the **Annotate** and **Classify** modules of
[spaCR](https://github.com/EinarOlafsson/spacr). It is the output of a Measure
run, so both modules can be exercised without segmenting or measuring anything
first.
## What is here
| Path | What it is |
|---|---|
| `data/` | 2,341 single-cell PNG crops, foldered by phenotype |
| `measurements.db` | The measurements, plus `png_list` and the annotation tables |
| `measurements/active_learning/` | The model card from the first annotation round |
| `settings/annotate_settings.csv` | Settings for the Annotate module |
| `settings/classify_settings.csv` | Settings for the Classify module |
## Labels: infected or not, by rule
`png_list.infected` holds a label for **every one of the 2,341 crops**, derived
from the measurements rather than by hand:
| Value | Meaning | The gesture that would produce it |
|---|---|---|
| `1` | no parasite in this cell | left click |
| `2` | at least one parasite | right click |
1,173 infected against 1,168 uninfected — 50.1%, which is about as balanced as
a binary training set gets without being resampled.
**How it was derived.** The `pathogen` table names the parent cell of every
detected parasite, so a cell is infected exactly when at least one pathogen row
points at it. That is a rule anyone can re-run and check, which a hand pass over
a subset is not.
**Independently cross-checked.** All 1,173 also have
`cell_pathogen_overlap_fraction > 0` — a column Measure computes by a different
route entirely. The two agree on every cell.
`annotate` is deliberately empty, so the module opens on a clean column and the
rule-based labels stay as a reference rather than as something to overwrite.
## Paths are RELATIVE, deliberately
Every path in `measurements.db` and in both settings files is relative to the
dataset root:
```
data/single_nucleus/uninfected/plate1_E01/cell_png/plate1_E01_19_1_12.png
```
A measurements database normally stores absolute paths, which name the machine
that made it and resolve nowhere else. spaCR's downloader rewrites these to
absolute on arrival, so the database works wherever it is unpacked. If you
unpack it by hand, do the same, or point spaCR at the folder and let it.
`<dataset>` in the settings files is a placeholder for the unpack location and
is substituted the same way.
## Plate layout
Four wells (E01, E02, L01, L02) from the plate published as
[`einarolafsson/spacr-example-measure`](https://huggingface.co/datasets/einarolafsson/spacr-example-measure),
which holds the merged arrays these crops were cut from.
## How spaCR downloads it
Everything here is also published as a single uncompressed **`spacr-example-annotate.tar`**, and
that is what spaCR fetches: one request instead of 2,365, a progress figure
that means something, and a download that stops when you press Cancel. It is
unpacked with tar's `data` filter, which refuses any member that would write
outside the destination folder.
The individual files are kept beside it so the set can be browsed and previewed
on this page. They are the same bytes; either is fine to use.
## Provenance
Produced by spaCR's Mask module and then its Measure module, from spaCR's own
example images, with `cpsam` for cells and nuclei and a Toxoplasma-specific
CPSAM checkpoint for pathogens. One field of fifty-two (`plate1_E02_20_1`)
failed to measure — a nucleus label spanning two cells — so the crops come from
51 fields. Absolute paths have been rewritten to relative throughout.
|