数字病理切片处理与研究

references/preprocessing.md

This reference describes PathML 3.0.8 stable. It corrects older examples that called nonexistent Pipeline.run(), omitted required mask/label names, or passed unsupported transform arguments.

Pipeline execution model

A Pipeline is an ordered list of Transform objects. Pipeline.apply(tile) modifies one Tile in place and returns it. A slide or dataset owns execution:

from pathml.core import HESlide
from pathml.preprocessing import BoxBlur, Pipeline, TissueDetectionHE

slide = HESlide("data/slide-001.svs", backend="openslide")
pipeline = Pipeline(
    [
        BoxBlur(kernel_size=5),
        TissueDetectionHE(mask_name="tissue"),
    ]
)

slide.run(
    pipeline,
    distributed=False,
    tile_size=512,
    tile_stride=512,
    level=0,
    tile_pad=False,
)

Stable facts:

• Pipeline(transform_sequence=None) and Pipeline.apply(tile) are public.
• Pipeline has no run() method.
• SlideData.run() and SlideDataset.run() apply pipelines.
• distributed=True is the default and may create a local Dask cluster using available cores. Start with distributed=False.
• Existing tiles are protected unless overwrite_existing_tiles=True.
• write_dir causes a <slide.name>.h5path write after processing.
• Pipeline.save() writes a pickle. A pickle is executable on load; never use a pipeline file from an untrusted source.

Stable transform imports

from pathml.preprocessing import (
    AdaptiveHistogramEqualization,
    BinaryThreshold,
    BoxBlur,
    CollapseRunsCODEX,
    CollapseRunsVectra,
    ForegroundDetection,
    GaussianBlur,
    HistogramEqualization,
    LabelArtifactTileHE,
    LabelWhiteSpaceHE,
    MedianBlur,
    MorphClose,
    MorphOpen,
    NucleusDetectionHE,
    Pipeline,
    QuantifyMIF,
    RescaleIntensity,
    SegmentMIF,
    SegmentMIFRemote,
    StainNormalizationHE,
    SuperpixelInterpolation,
    TissueDetectionHE,
)

There is no stable transform='...' registry and no safe reason to construct transforms from arbitrary Python expressions. Parse a strict allowlisted config, then instantiate known classes explicitly.

Tissue detection

from pathml.preprocessing import TissueDetectionHE

tissue = TissueDetectionHE(
    mask_name="tissue",
    use_saturation=True,
    blur_ksize=17,
    threshold=None,          # Otsu when None
    morph_n_iter=3,
    morph_k_size=7,
    min_region_size=5000,
    max_hole_size=1500,
    outer_contours_only=False,
)

The transform expects an H&E uint8 tile. It:

1. uses HSV saturation or greyscale;
2. median-blurs;
3. applies Otsu or the explicit threshold;
4. performs morphological opening and closing;
5. keeps foreground regions under the configured area/hole policy; and
6. writes tile.masks["tissue"].

mask_name is required in practice; None fails when apply() runs. min_region_size, max_hole_size, and morphology kernels are measured in pixels at the processing level. Re-tune if level or MPP changes.

Tissue detection is tile-local. It can disagree at tile edges, and PathML does not automatically remove background tiles from the pipeline. Compute and record tissue coverage after the mask exists:

coverage = float((tile.masks["tissue"] > 0).mean())
keep = coverage >= 0.50

Choose the coverage rule on training data and preserve rejected-tile counts.

Whitespace and artifact QC

These transforms write tile labels, not pixel masks:

from pathml.preprocessing import LabelArtifactTileHE, LabelWhiteSpaceHE

whitespace = LabelWhiteSpaceHE(
    label_name="mostly_white",
    greyscale_threshold=230,
    proportion_threshold=0.5,
)
artifact = LabelArtifactTileHE(label_name="artifact")

• LabelWhiteSpaceHE labels a tile when the proportion of greyscale pixels above the threshold exceeds proportion_threshold.
• LabelArtifactTileHE is a fixed rule-based HSI heuristic for whitespace, dark regions, and pen-like colors. It exposes only label_name; older examples with pen_threshold or bubble_threshold are invalid.

Neither is a complete slide-quality system. Review representative overlays and track blur, folds, bubbles, pen, tissue coverage, clipping, color drift, missing channels, and focus separately. Do not convert a heuristic QC flag into a clinical quality judgment.

The dependency-free helper provides a deliberately simple synthetic/local check, not a replacement for PathML:

python scripts/image_qc.py synthetic --width 256 --height 256
python scripts/image_qc.py inspect --image tests/fixtures/synthetic.ppm --root .

It reports brightness/saturation and a coarse tissue-like mask using bounded pixels. PNG/JPEG/TIFF input needs Pillow, imported only after argument validation.

H&E stain normalization and separation

from pathml.preprocessing import StainNormalizationHE

normalizer = StainNormalizationHE(
    target="normalize",             # normalize | hematoxylin | eosin
    stain_estimation_method="macenko",  # macenko | vahadane
    optical_density_threshold=0.15,
    regularizer=0.1,
    angular_percentile=0.01,
    background_intensity=245,
)

Stable StainNormalizationHE does not accept tissue_mask_name, target_od, or target_concentrations. It accepts stain_matrix_target_od and max_c_target, and supplies fixed defaults.

To fit a reference:

normalizer.fit_to_reference(training_reference_rgb)
normalized_rgb = normalizer.F(source_rgb)

Leakage controls:

• Select the reference and tune OD parameters using training slides only.
• Freeze the fitted stain matrix and target concentration before validation/test.
• Do not choose a reference because test performance looks better.
• Record source slide pseudonym, region coordinates, level, MPP, method, and all fitted arrays without direct identifiers.
• Fit on tissue-rich, artifact-free RGB regions. Stable API does not consume a tissue mask directly, so crop/filter the reference beforehand.

Macenko and Vahadane are model-based color standardization methods, not guarantees that biological staining becomes comparable. Preserve raw inputs and assess whether normalization removes task-relevant signal or amplifies artifacts.

Simple H&E nucleus mask

from pathml.preprocessing import NucleusDetectionHE

nuclei = NucleusDetectionHE(
    mask_name="nuclei",
    stain_estimation_method="vahadane",
    superpixel_region_size=10,
    n_iter=30,
)

This is a simple transform: hematoxylin separation, superpixel interpolation, and Otsu thresholding. It writes a binary tile mask. It is not HoVer-Net, does not assign nucleus classes, and should not be treated as a validated cell count. Inspect touching objects, fragments, necrosis, stain failure, and tile boundaries.

Binary and morphology building blocks

Useful stable signatures:

from pathml.preprocessing import BinaryThreshold, MorphClose, MorphOpen

threshold = BinaryThreshold(
    mask_name="foreground",
    use_otsu=True,
    threshold=0,
    inverse=False,
)
opened = MorphOpen(mask_name="foreground", kernel_size=5, n_iterations=1)
closed = MorphClose(mask_name="foreground", kernel_size=5, n_iterations=1)

BinaryThreshold.apply() creates a named mask. MorphOpen and MorphClose modify the named mask. Match dtype, polarity, and dimensions explicitly.

Recommended pilot pipeline

from pathml.preprocessing import (
    BoxBlur,
    LabelArtifactTileHE,
    LabelWhiteSpaceHE,
    Pipeline,
    StainNormalizationHE,
    TissueDetectionHE,
)

pipeline = Pipeline(
    [
        LabelWhiteSpaceHE(
            label_name="mostly_white",
            greyscale_threshold=230,
            proportion_threshold=0.8,
        ),
        LabelArtifactTileHE(label_name="artifact"),
        BoxBlur(kernel_size=3),
        TissueDetectionHE(
            mask_name="tissue",
            min_region_size=5000,
            outer_contours_only=False,
        ),
        StainNormalizationHE(
            target="normalize",
            stain_estimation_method="macenko",
        ),
    ]
)

QC labels do not short-circuit later transforms. If expensive stages should run only on accepted tiles, write an explicit custom transform or bounded manual loop with a documented policy. Keep the logic deterministic and test it.

Bounded dry run

PathML has no max_tiles argument. Use islice for a local pilot:

from itertools import islice

sampled = []
for tile in islice(
    slide.generate_tiles(shape=512, stride=512, pad=False, level=0),
    16,
):
    pipeline.apply(tile)
    sampled.append(
        {
            "coords_ij": tile.coords,
            "shape": tuple(tile.image.shape),
            "tissue_fraction": float((tile.masks["tissue"] > 0).mean()),
            "mostly_white": bool(tile.labels["mostly_white"]),
            "artifact": bool(tile.labels["artifact"]),
        }
    )

Plan first:

python scripts/plan_pipeline.py \
  --width 100000 --height 80000 \
  --tile-size 512 --stride 512 \
  --pipeline TissueDetectionHE,LabelWhiteSpaceHE,StainNormalizationHE \
  --max-tiles 1000000

The planner never opens the slide or imports PathML. Supply dimensions from a trusted technical metadata inspection. Its storage estimate uses h5path's float16 image/mask payload, excludes counts/labels/metadata, and assumes the supplied channel count remains unchanged. It is not peak RAM or a guarantee of available disk. At nonzero levels supply exact --level-width/--level-height.

Masks, labels, padding, and overlap

• A tile mask's first two dimensions must match the tile image.
• Use distinct semantic names (tissue, nuclei, cell_segmentation) and record whether each mask is binary, semantic, or instance-labeled.
• Instance masks use 0 for background and positive integer object IDs.
• tile_pad=True introduces zeros. Document whether padded pixels are ignored in QC, stain fitting, loss, and stitching.
• Stable SlideData.generate_tiles() cannot slice slide-level masks into padded tiles.
• Overlap duplicates tissue/cells. Deduplicate by slide coordinates or use an explicit blending/cropping policy before counting.
• Keep masks at the same level as their coordinates. Resampling an instance mask requires nearest-neighbor interpolation and relabel/QC.

Train/validation/test leakage checklist

Do all splitting before:

• choosing stain references;
• estimating QC or tissue thresholds;
• fitting feature scalers;
• learning augmentations or color distributions;
• selecting segmentation parameters;
• extracting overlapping tiles;
• constructing graphs; or
• calibrating model thresholds.

All tiles, regions, serial sections, and repeat scans from one patient belong to one split. If site/scanner generalization is the target, reserve entire sites or scanners as designed. Record exclusions before viewing test outcomes.

Reproducibility record

For each run, retain:

• PathML and dependency lock versions;
• source SHA-256 and pseudonymous IDs;
• backend, level, downsample, MPP, tile size/stride/pad;
• ordered transform names and all constructor values;
• fitted stain arrays and training-only reference provenance;
• QC/mask definitions and per-slide accept/reject totals;
• Dask configuration, worker count, CPU/GPU, and failure/retry policy;
• code revision, random seeds, split manifest hash, and output hashes.

Sources and further reading

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

• Stable pipeline guide: https://pathml.readthedocs.io/en/stable/creating_pipelines.html
• Stable execution guide: https://pathml.readthedocs.io/en/stable/running_pipelines.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 pipeline source: https://github.com/Dana-Farber-AIOS/pathml/blob/v3.0.8/pathml/preprocessing/pipeline.py
• Macenko et al. (2009): https://doi.org/10.1109/ISBI.2009.5193250
• Vahadane et al. (2016): https://doi.org/10.1109/TMI.2016.2529665