Digital Pathology: a pipeline architecture for trustworthy imaging.

01 · Pipeline Engineering

Turn an experimental method into an executable system.

Microscopy analysis often begins as a working method understood by one scientist: instrument settings, processing decisions, and scripts developed around a particular experiment. Fovea Lab provides a structured environment for turning that method into a pipeline that can be run, examined, and developed over time.

A scientist can begin in the visual workspace, ask the agent to help develop a stage, or open the underlying code when direct control is needed. Each way of working modifies the same pipeline rather than creating a separate version of the analysis. The scientific purpose of the experiment remains connected to the computation, allowing the workflow to mature from early exploration into a system that another researcher can understand and use.

02 · Evidence Integrity

Make confidence part of the computation.

A final measurement does not reveal whether the source data were complete, whether a model behaved unexpectedly, or whether an important parameter changed between runs. Fovea Lab evaluates the analysis as it runs and preserves the evidence needed to understand the result afterward.

Checks are placed at the stages where failures can occur. Scientists can return from a result to the pipeline that produced it, inspect whether the run remained within accepted conditions, and trace unexpected behavior to the point where it entered the analysis. Verification becomes part of the workflow itself rather than a report reconstructed after the experiment is finished.

03 · Agentic Development

Let AI improve the workflow without taking control away from the scientist.

Fovea Lab gives agents access to the structured experimental context that governs the pipeline. Instead of producing isolated code without understanding the surrounding system, an agent can inspect a specific stage, propose a bounded revision, and evaluate its effect against criteria established by the scientist.

A proposed change remains separate from the approved workflow until it has been reviewed. Its effect can be compared with the current version, while the decision to adopt it remains with an authorized user. AI can accelerate pipeline development while the platform preserves accountability and a stable record of what was actually run.

Pipeline architectures, module by module.

Pipeline backbone — reshapes per modality. Select a module for its sub-steps, tools, and deliverable.

Digital Pathology

High confidence

Ingest

Read gigapixel pyramidal whole-slide images from vendor-specific formats (SVS, NDPI, MRXS, DICOM WSI, OME-TIFF), extract scanner and acquisition metadata, convert to chunked cloud-native representations, and generate a coordinate-indexed tile inventory that is analysis-ready. This stage is the foundation for all downstream QC, preprocessing and inference — correct tiling parameters and MPP calibration here prevent hard-to-detect spatial errors later.

Operations

  1. Format detection and validationDetect vendor format (Aperio SVS, Hamamatsu NDPI, 3DHISTECH MRXS, Leica SCN, DICOM WSI, OME-TIFF) and validate file integrity before opening.
  2. Pyramidal WSI loadingOpen the full image pyramid, enumerate resolution levels, verify level dimensions, and extract embedded scanner metadata (MPP, objective, ICC profile).
  3. Low-resolution tissue previewRender thumbnail at 1.25x or 2.5x level; compute foreground mask via Otsu thresholding on the LAB luminance channel to estimate tissue coverage.
  4. Tiling with overlapGenerate coordinate-indexed tiles at the chosen analysis magnification (typically 20x or 40x) with configurable stride and overlap; record tile coordinates in physical microns.
  5. Conversion to cloud-native formatoptionalWrite image pyramid to OME-Zarr or OME-TIFF with chunked storage; optionally convert to DICOM WSI for clinical interoperability.
  6. Manifest creationPersist slide, level and tile inventory (slide_id, MPP, objective, tile coordinates, pyramid levels) to Parquet or JSON.

Inputs

  • Raw whole-slide images in vendor pyramidal formatsSVS · NDPI · MRXS · SCN · OME-TIFF · DICOM WSI · BIF · QPTIFF

Outputs

  • Cloud-native tiled WSI pyramidOME-Zarr · OME-TIFF · DICOM WSI
  • Tile coordinate manifestParquet · JSON
  • Foreground tissue maskPNG · OME-Zarr

QC checks

  • MPP calibration plausibility — Check that MPP is within expected range for stated magnification (e.g. 0.20–0.35 µm/px at 40x). Flag mismatches that indicate scanner metadata corruption.
  • Pyramid completeness — Verify all expected resolution levels are present and readable; flag truncated pyramids.
  • Tile integrity sampling — Randomly sample 1% of tiles and verify JPEG/JP2 decode success; detect partial file transfers or corruption.
  • Foreground fraction floor — Flag slides with <10% tissue coverage as likely empty or grossly under-sectioned.
  • Format / version compatibility — Confirm OpenSlide or chosen reader can parse the format without errors; log vendor format version.

Failure modes it guards against

  • MPP absent or incorrect in vendor metadata — Cross-validate MPP against known tissue structures (e.g. red blood cell ~7 µm); fall back to calibration table by scanner model.
  • Incomplete file transfer / truncated WSI — Checksum (MD5/SHA256) the source file before ingest; validate all pyramid levels open.
  • Proprietary format not supported by reader — Maintain format compatibility matrix; fall back to Bio-Formats or HALO for edge cases.
  • High tile overlap causing memory/storage explosion — Set overlap only as large as needed for the downstream model's receptive field.
  • Scale/resolution confusion between pyramid levels — Record and validate analysis_level and corresponding MPP in manifest; assert before inference.

Tools (17)

  • OpenSlidePRIMARYOSS · LGPL-2.1

    Read vendor whole-slide image formats (SVS, NDPI, MRXS, SCN, BIF, QPTIFF) via a unified C library with Python bindings.

  • tifffilePRIMARYLibrary · BSD-3-Clause

    Read/write TIFF, BigTIFF, OME-TIFF and pyramidal TIFF variants; used for OME-TIFF I/O and low-level tile access.

  • ome-zarr-pyPRIMARYOSS · BSD-2-Clause

    Read/write OME-Zarr (NGFF) compliant multiscale image arrays; used to write analysis-ready pyramids to object storage.

  • tiffslideOSS · BSD-3-Clause

    Cloud-native drop-in replacement for openslide-python built on tifffile; reads WSI from object storage without local download.

  • cucimOSS · Apache-2.0

    GPU-accelerated WSI I/O (TIFF/SVS), NVJPEG decoding, and pathology image primitives for high-throughput tiling pipelines.

  • Bio-FormatsOSS · GPL-2.0 (open) / commercial (OME)

    Java library (with Python wrapper) supporting 150+ microscopy and pathology formats including Zeiss CZI, Leica LIF, and others beyond OpenSlide's scope.

  • pyvipsLibrary · LGPL-2.1

    Fast image processing library for large-image tiling, format conversion and pyramidal TIFF construction using libvips.

  • large-imageOSS · Apache-2.0

    Python library for serving and tiling large images including WSI; integrates OpenSlide, tifffile, DICOM and GDAL sources.

  • pydicomOSS · MIT

    Read DICOM WSI files and extract DICOM tags (pixel data, UIDs, scanner metadata).

  • wsidicomOSS · Apache-2.0

    High-level Python library specifically for DICOM WSI reading, tiling and writing; supports DICOM VL Whole Slide Microscopy Image.

  • highdicomOSS · MIT

    Python library for reading/writing DICOM Structured Reports (SR), Segmentation objects, and WSI; used for DICOM-compliant output.

  • TIAToolboxOSS · BSD-3-Clause

    End-to-end computational pathology toolbox with WSI reading, tiling, model inference and visualization.

  • QuPathOSS · GPL-3.0

    Cross-platform WSI viewer and analysis platform with scripting support (Groovy/Python); used for QC, annotation and tile export.

  • PathMLOSS · Apache-2.0

    Digital pathology preprocessing and ML pipeline toolkit; wraps OpenSlide and stain normalization for rapid pipeline construction.

  • MONAIOSS · Apache-2.0

    Medical imaging AI framework with MONAI Deploy for clinical WSI workflows and MONAI Label for interactive annotation.

  • Indica Labs HALOCommercial · commercial

    Commercial WSI analysis platform supporting 30+ scanner formats; used when vendor-certified format reading is required.

  • VisiopharmCommercial · commercial

    Commercial WSI analysis platform with certified scanner interfaces and regulatory-grade traceability.

Metadata captured

FieldTypeExampleStandard
pipeline_run_idstringdp-run-2026-06-06-a1b2c3
software_versionsobject{'openslide': '4.0.0', 'tifffile': '2024.1.1', 'ome-zarr-py': '0.9.0', 'python': '3.11', 'container': 'sha256:abc'}REMBI
slide_idstringTCGA-A1-A0SD-01Z-00-DX1DICOM WSI
scanner_make_modelstringLeica Aperio AT2DICOM WSI
objective_magnificationfloat (x)40.0DICOM WSI
mppfloat (micrometer_per_pixel)0.2425DICOM WSI
pyramid_levelsint (count)5OME-NGFF
tile_size_pxint (pixel)512
tile_overlap_pxint (pixel)64
compression_codecenumJPEGDICOM WSI
icc_color_profileuris3://scanner-profiles/aperio-at2-sRGB.iccDICOM WSI
slide_dimensions_pxarray (pixel)98816, 75264OME-NGFF
foreground_fractionfloat (fraction)0.68
n_tiles_totalint (count)12480
output_formatenumome-zarrOME-NGFF
input_dataset_uriuris3://client-study/raw/slide01.svs
output_dataset_uriuris3://client-study/derived/slide01.zarrOME-NGFF

You get

  • Cloud-native WSI pyramid · OME-Zarr
  • Tile coordinate manifest · Parquet
  • Slide ingest metadata record · JSON

References