数字病理切片处理与研究

references/multiparametric.md

This reference targets PathML 3.0.8 stable. It distinguishes generic loading support from a dedicated analysis implementation and corrects older examples that invented MERFISH decoders, spectral unmixing options, named-channel arguments, or a DeepCell cloud prediction endpoint.

Stable slide types and array order

from pathml.core import CODEXSlide, MultiparametricSlide, VectraSlide

codex = CODEXSlide(
    "data/codex_region.tif",
    name="slide-001",
    backend="bioformats",
)
vectra = VectraSlide(
    "data/vectra_component_data.tif",
    name="slide-002",
)
generic = MultiparametricSlide(
    "data/multiplex.ome.tiff",
    name="slide-003",
)

Bio-Formats returns PathML arrays as:

(i, j, z, c, t) = (row, column, z-plane, channel, time/cycle)

Generic Bio-Formats support means a format may be readable. It does not mean PathML implements registration, decoding, spectral unmixing, compensation, autofluorescence correction, cell phenotyping, or platform-specific QC for that format.

Stable PathML has convenience classes for CODEX and Vectra. It can load examples of MERFISH/Visium-like image data through generic backends, but it has no MERFISHSlide, DecodeMERFISH, or AssignTranscripts stable API.

Channel manifest

Create a local, versioned channel manifest before processing:

channel_index,channel_name,marker,cycle,z_plane,role,exposure,batch
0,DAPI,DAPI,0,2,nuclear,100,run-01
1,CD45,CD45,0,2,membrane,150,run-01
2,CD3,CD3,1,2,measurement,200,run-01

Validate:

• channel index is zero-based, unique, and matches array order;
• DAPI/nuclear and membrane/cytoplasm choices are biologically and technically appropriate;
• cycle and Z-plane mappings are documented;
• blank, autofluorescence, isotype, and positive controls are identified;
• saturation, clipping, hot pixels, bleed-through, exposure, registration, and missing channels are assessed;
• marker names are not inferred from position alone.

Do not include patient names, accession numbers, or other direct identifiers.

CODEX collapse

Stable signature:

from pathml.preprocessing import CollapseRunsCODEX, Pipeline

pipeline = Pipeline([CollapseRunsCODEX(z=2)])

CollapseRunsCODEX(z) expects (i, j, z, c, t), combines c and t into one channel axis, selects the zero-based Z-plane, and produces (i, j, c*t).

It does not:

• register cycles;
• choose a focal plane automatically;
• subtract background;
• reorder markers from a manifest;
• aggregate with max/mean/median; or
• correct illumination or bleed-through.

Perform and validate those operations upstream with a protocol appropriate to the acquisition system. Record the exact flattened (cycle, channel) → output index mapping.

The actual C-order reshape maps (channel, cycle) to output_index = channel * n_cycles + cycle (cycle varies fastest). Verify it with a channel-coded synthetic array; do not assume cycle-major order. The released Z-index assertion checks the flattened channel axis by mistake, so validate 0 <= z < image.shape[2] yourself before calling it.

Vectra collapse

Stable signature:

from pathml.preprocessing import CollapseRunsVectra

collapse = CollapseRunsVectra()

It applies numpy.squeeze to coerce the image toward (i, j, c). It does not accept wavelengths and does not perform spectral unmixing or autofluorescence correction. Supply already unmixed component data or perform those steps with validated upstream software. Verify that squeezing singleton axes did not remove an axis whose semantics must be retained.

Segmentation choices

Preferred no-network posture

For sensitive images, use an institution-approved local segmenter with a reviewed, checksummed local model and network disabled. PathML's generic local ONNX Inference can run compatible models, but Mesmer-specific pre/postprocessing must match the model card exactly.

PathML 3.0.8 has no fully offline, nondeprecated Mesmer convenience class that accepts a pre-provisioned model without trying a network download. Plan this constraint before choosing PathML's Mesmer wrapper.

SegmentMIFRemote: local inference after a model download

Stable signature:

SegmentMIFRemote(
    model_path="temp.onnx",
    nuclear_channel=<integer index>,
    cytoplasm_channel=<integer index>,
    image_resolution=0.5,
    preprocess_kwargs=None,
    postprocess_kwargs_nuclear=None,
    postprocess_kwargs_whole_cell=None,
)

Despite the name, v3.0.8 does not send images to a DeepCell prediction service. At construction it performs an HTTP GET from:

https://huggingface.co/pathml/test/resolve/main/mesmer.onnx

It writes the response to temp.onnx, loads that ONNX model, and runs pixels locally with ONNX Runtime. The outbound request discloses ordinary connection metadata (for example IP address and request headers) to Hugging Face; image pixels, channel data, and PathML metadata are not uploaded by this stable source.

Security and reproducibility limitations:

• construction has a network side effect;
• there is no built-in checksum, signature, offline flag, timeout, or size cap;
• inherited local=False forces temp.onnx, even if SegmentMIFRemote receives a different model_path; a supplied destination does not prevent overwriting that working-directory file;
• the model supports 256×256 input and 0.5 µm/pixel in stable code;
• nuclear_channel and cytoplasm_channel are integer indices, not names.

Validate HWC input (despite a contradictory channel-first wrapper docstring), both spatial dimensions equal to 256, distinct valid channel indices, and 0.5 µm/pixel before constructing/running the wrapper. The source only prints for other resolutions and its shape check uses and, so one incorrect spatial axis can escape that check; it does not resample to the model resolution for you.

Do not instantiate it until the user explicitly consents to that endpoint and download. Do not use it in a network-disabled workflow.

On 2026-10-01 the public Hugging Face model metadata listed both artifacts and HEAD requests followed download redirects successfully (Mesmer 104,218,552 bytes; HoVer-Net test model 151,068,601 bytes). These were unauthenticated metadata/HEAD probes only, not weight downloads, checksum verification, or inference tests.

Deprecated SegmentMIF

SegmentMIF imports deepcell.applications.Mesmer and calls its prediction API locally. PathML emits a deprecation warning directing users to SegmentMIFRemote. deepcell is not a PathML extra and is not declared in PathML's PyPI dependencies. Depending on the DeepCell version/cache, model initialization may fetch weights.

Do not silently install an unpinned DeepCell stack or assume compatibility with PathML's pinned Python/TensorFlow ecosystem. If legacy replication requires it, lock the complete environment, pre-provision and verify artifacts, test with synthetic data, disable network during sensitive runs, and document the deprecation.

Explicit network consent template

Before any download or hosted inference, present:

Action: download model | download public dataset | upload image for prediction
Destination: exact HTTPS host and path
Outbound data: exact pixels/channels/metadata/identifiers, or "none; GET only"
Inbound artifact: name, expected bytes, version, SHA-256/signature
Local destination: approved path
Retention/logging: vendor and institutional policy
Authorization: data-use agreement/consent/waiver and user approval
Alternative: local, network-disabled method

Only proceed after explicit opt-in. A future hosted endpoint that receives image data needs a new disclosure; do not infer permission from consent to download a model. Never upload PHI by default.

Quantification

Read Bio-Formats tiles with normalize=False for measurement. Its default normalization casts to 8-bit and can destroy quantitative dynamic range; h5path then separately stores images/masks as float16. Keep authoritative raw channels and integer instance maps outside that lossy storage path.

Stable transform:

from pathml.preprocessing import QuantifyMIF

quantify = QuantifyMIF(segmentation_mask="cell_segmentation")

Input requirements:

• tile image (i, j, channels);
• instance segmentation (i, j) or (i, j, 1);
• 0 is background and each object has a unique positive integer;
• tile slide type has fluorescence stain;
• tile.coords is present.

QuantifyMIF.apply(tile) writes an AnnData object to tile.counts. Stable output contains:

• X: per-object mean intensity for each channel;
• layers["min_intensity"] and layers["max_intensity"];
• obs["label"], filled_area, euler_number, x, and y;
• obsm["spatial"]: (x, y) centroids.

Channel variables are generated from numeric positions. Assign marker names only after matching the validated channel manifest:

counts = tile.counts
assert counts.n_vars == len(channel_names)
counts.var_names = channel_names

Do not interpret raw intensity as abundance without validated background, normalization, exposure, compensation/unmixing, segmentation, and batch policies.

Coordinate conversion

Tile.coords uses selected-level (i, j). QuantifyMIF adds tile offsets to region centroids, then stores:

• obs["y"]: row coordinate;
• obs["x"]: column coordinate;
• obsm["spatial"]: [x, y].

Convert selected-level pixels to level-0 and physical units:

x_um = x_selected_level * level_downsample * mpp_x
y_um = y_selected_level * level_downsample * mpp_y

Record whether coordinates are pixel centers, integer-rounded centroids, or continuous region centroids. Preserve source level and MPP. Do not combine slides with different resolutions in a shared coordinate space without conversion. The scalar-downsample formula assumes a validated series/pyramid transform and correct region origins; Bio-Formats series need not form a pyramid, and PathML's OpenSlide fractional-downsample origin issue must be excluded first.

Spatial/multiplex schema

For exchange, use one row per cell:

cell_id,patient_id,slide_id,tile_i,tile_j,x,y,coordinate_unit,level,
segmentation_label,area,marker:DAPI,marker:CD3,marker:CD8,...

Required invariants:

• (slide_id, cell_id) is unique;
• coordinates are finite and nonnegative;
• coordinate_unit is explicit (level_pixels, level0_pixels, or um);
• all rows for a slide agree on level/unit/channel schema;
• marker values are finite or use one documented missing-value policy;
• instance IDs do not collide when tiles are merged;
• overlapping-tile duplicate cells are reconciled before counts;
• patient/slide split is included or joinable without PHI.

Validate bounded local CSV:

python scripts/validate_spatial_schema.py multiplex \
  --input derived/cells.csv \
  --root . \
  --marker-columns marker:DAPI,marker:CD3,marker:CD8

Segmentation and marker QC

Review representative training regions:

• nuclear and whole-cell overlays;
• object count, area, eccentricity, border-touching fraction, holes/fragments;
• merge/split errors and compartment consistency;
• negative/blank controls and background distributions;
• per-channel saturation and dynamic range;
• spatial striping, illumination, cycle registration, and tissue folds;
• cell density by tissue compartment and slide;
• marker distributions by batch/site/scanner;
• effects of image resolution and channel choice.

Freeze QC thresholds before the test set. Report exclusions and sensitivity analyses rather than hiding failed tiles.

Combining AnnData safely

Before concatenation:

• make obs_names globally unique with pseudonymous slide and cell IDs;
• store patient_id, slide_id, site, batch, and split in obs;
• enforce identical marker identity/order or perform an explicit outer join;
• record raw versus transformed layers;
• preserve coordinate unit/level per slide;
• never batch-correct using held-out test data;
• keep spatial neighbors within slides unless a cross-slide graph is scientifically defined.

Threshold cell typing from markers is a research annotation procedure, not a diagnosis. Use controls and domain review, and report ambiguous/unassigned cells.

Stable versus absent APIs

Present in stable:

• MultiparametricSlide, VectraSlide, CODEXSlide
• CollapseRunsCODEX(z)
• CollapseRunsVectra()
• SegmentMIFRemote(...)
• deprecated SegmentMIF(...)
• QuantifyMIF(segmentation_mask)

Not present as stable APIs:

• named-channel arguments such as "DAPI" or "CD45" for segmentation;
• CollapseRunsCODEX(method=..., background_subtract=...);
• CollapseRunsVectra(wavelengths=..., unmix=True);
• MERFISHSlide, DecodeMERFISH, or AssignTranscripts;
• a DeepCell image-upload URL in PathML;
• automatic marker names in QuantifyMIF;
• automatic cell-type annotation.

Sources and further reading

API baseline reviewed 2026-10-01 using the released wheel/tag; hosted docs may lag.

• Stable loading guide: https://pathml.readthedocs.io/en/stable/loading_slides.html
• Stable preprocessing API: https://pathml.readthedocs.io/en/stable/api_preprocessing_reference.html
• Stable transforms source: https://github.com/Dana-Farber-AIOS/pathml/blob/v3.0.8/pathml/preprocessing/transforms.py
• Stable inference source: https://github.com/Dana-Farber-AIOS/pathml/blob/v3.0.8/pathml/inference/inference.py
• Stable multiplex tutorial: https://pathml.readthedocs.io/en/stable/examples/link_multiplex_if.html
• Stable CODEX tutorial: https://pathml.readthedocs.io/en/stable/examples/link_codex.html
• Greenwald et al. (2022), Mesmer: https://doi.org/10.1038/s41587-021-01094-0
• Omar et al. (2025), antibody-based multiplex workflows: https://doi.org/10.1016/j.labinv.2025.104220