cellprofiler – skill Thiết kế & UI cho agent AI

CellProfiler quantitative microscopy

Use this skill when a user needs a repeatable CellProfiler .cppipe, nuclear counts, nuclear

fluorescence, or batch microscopy measurements. The bundled assay accepts **one 2D grayscale

TIFF nuclear channel per field**, with black-is-zero (MINISBLACK) pixels and bright nuclei on a

dark background. Palette and white-is-zero TIFFs need an explicit conversion. For volumetric

segmentation, multichannel cell painting, or tissue-specific models, design a separate pipeline

and validate those assumptions rather than silently projecting or splitting the images.

The official application and manual remain 4.2.8. PyPI publishes 4.2.8.1; its seven

modules used here and embedded Threshold module match the 4.2.8 source, but this review did

not execute that native distribution. Keep the helper environment separate from CellProfiler's

older dependency stack; see the runtime reference for the verification boundary.

Workflow

1. Establish the acquisition unit: plate, well, site, time point if present, pixel size, nuclear

channel identity, camera bit depth, exposure, and biological replicate. Keep original image

intensities. Convert proprietary formats explicitly with Bio-Formats before using this helper.

2. Create the CSV manifest below. image_path is absolute or relative to the manifest; sample IDs use letters, digits, dots, dashes, or underscores; sample IDs

and plate/well/site combinations are unique. Use a nonnumeric sample ID such as sample_001:

LoadData infers column types and can otherwise turn 001 into 1. Avoid surrounding

whitespace in identifiers. TIFFs must be uint8 or uint16, single plane/series/resolution, and

nonconstant. The helper rejects RGB, z-stacks, and float images rather than guessing channels.

3. Use [assets/nuclei.cppipe](assets/nuclei.cppipe) as a starting pipeline: LoadData →

IdentifyPrimaryObjects → intensity/size measurements → outline overlay → CSV export.

The initial diameter range is 8–80 pixels, with global Otsu thresholding, no threshold

smoothing, and border objects excluded. Calibrate this

range from representative images and acquisition pixel size before comparing conditions.

4. Run a small pilot spanning controls, low/high density, dim images, and plate edges. Inspect

saved overlays for missed nuclei, splits, merges, and edge exclusions. Adjust thresholding

and declumping in CellProfiler, export the tuned .cppipe, and pass --pipeline to preserve

it. Do not choose settings separately for each treatment to make their counts agree.

5. Freeze the tuned pipeline and analyze the batch. Review input saturation warnings, zero

counts, count/area distributions, and control behavior. Aggregation for inference belongs at

the biological replicate level; thousands of cells from one well are not independent wells.

Run the bounded assay

From this skill directory, create images.csv:

sample_id,image_path,plate,well,site
control_A01_1,images/control_A01_1_DAPI.tif,Plate1,A01,1

python scripts/nuclei_assay.py prepare images.csv load_data.csv
python scripts/nuclei_assay.py run images.csv results --executable cellprofiler
python scripts/nuclei_assay.py summarize results

run requires a fresh/empty output directory and executes CellProfiler with -c -r, a saved

pipeline copy, --data-file, output folder, and --done-file. Success requires exit code zero,

a Complete marker, and valid measurement tables. It records the command, pipeline checksum,

input image checksums, and sample QC in assay_qc.json before execution, retaining failed

status and the error if execution or output validation fails. CellProfiler output goes to

cellprofiler.log. Rerun in a new output folder. summarize checks CSV contents independently;

it does not prove an engine run completed.

Custom pipelines must preserve DNA, Nuclei, Metadata_Sample, integer-dtype scaling, and

the unprefixed single-object Image.csv/Nuclei.csv export contract. Keep the required

intensity/area measurements. A renamed object set or different intensity scale needs a

corresponding helper adaptation, not an unchecked --pipeline substitution.

The executable can also be the CellProfiler application launcher or a local container launcher;

see [references/runtime-and-qc.md](references/runtime-and-qc.md) for the container target,

filesystem mapping, and verification evidence. prepare and summarize work without CellProfiler.

Interpret the outputs

• Image.csv: one image/field row, including Count_Nuclei and acquisition metadata.
• Nuclei.csv: one accepted object per row, with mean/integrated DNA intensity, area, and shape.
• *_nuclei.png: green nuclear boundaries over the input image for visual QC.
• pipeline.cppipe and cellprofiler.done: the exact pipeline copy and engine completion marker.
• assay_qc.json: run status, unique image/object keys, exact counts, finite mean/integrated

intensity and positive area checks, field mean area in pixels, and storage saturation flags.

LoadData ignores camera metadata for scaling in this asset and divides by the integer storage

maximum: uint8 → 255, uint16 → 65535. A 12-bit camera stored in uint16 therefore has a maximum

near 0.0625. Do not compare intensities across different bit depths, exposures, gains, or staining

batches without an explicit calibration. A saturated image can pass segmentation while its

intensity measurement is unusable. Illumination correction and background subtraction are

assay-specific additions; this starter does neither.

The saturation fraction only counts pixels at the storage maximum. A 12-bit detector may

saturate at 4095 while the uint16 storage maximum is 65535; inspect the known acquisition ceiling

separately. Integrated intensity sums pixel values and may exceed 1; only per-pixel mean

intensity is constrained to 0–1. The field's mean nuclear intensity weights each nucleus

equally, rather than weighting each pixel equally.

A count check cannot prove correct segmentation. Inspect overlays and independently annotated

fields; report boundary exclusions and segmentation errors alongside the biological result.

The optional synthetic engine test targets a known three-nucleus example, not assay performance

on unseen cell types. It was skipped in the current review because no engine was configured.

Sources

• [Official example pipelines](https://cellprofiler.org/examples): choose an assay-specific starting point.
• [CellProfiler 4.2.8 manual](https://cellprofiler-manual.s3.amazonaws.com/CellProfiler-4.2.8/index.html): module settings and interpretation.
• [Headless batch processing](https://cellprofiler-manual.s3.amazonaws.com/CellProfiler-4.2.8/help/other_batch.html): command-line execution.
• [Current application download](https://cellprofiler.org/releases) and [PyPI distribution](https://pypi.org/project/cellprofiler/): distinct release targets.