Whole Slide Imaging
The digitised glass slide — a very large, multi-resolution image file that is the substrate for every computational pathology method and every operational bottleneck in the digital workflow.
What it is
A scanner images a slide at high magnification and stores the result as a pyramid: the full resolution level plus progressively downsampled copies, so a viewer can pan and zoom without loading the whole thing. Files run to gigabytes each. Formats are vendor-specific — SVS (Aperio/Leica), NDPI (Hamamatsu), plus DICOM as the interoperability standard — which is why the group maintains an imageTypes repo and why format handling shows up as its own concern.
Three properties drive nearly everything downstream. Magnification (commonly 20x or 40x scanning, with analysis often at 10x or 20x) sets what any model can physically resolve. Pyramid structure means “which level am I reading” is a real analysis parameter, not an implementation detail. Embedded metadata carries scanner type, resolution, and — critically — label images that may contain patient identifiers, which is why de-identification is not just a matter of renaming files.
Because files are large, the operational layer is a genuine research subject in its own right: scanning throughput, transfer reliability, and storage cost are not incidental.
Why it matters for my work
WSI sits underneath three separate strands of the group’s work simultaneously.
As a research substrate: QuPath Tooling and every AI project consume WSIs; the de-identified image cohort lives on the Memorial share and explicitly not in git.
As an operational subject: Scanning Time in Real Life studies scanner utilisation, transfer, and turnaround across Leica scanners and Sectra integration — the file size is the reason those bottlenecks exist.
As a governance problem: the WSI transfer SOP (v1.2) covers local-first scanning, monitored transfer to Sectra, watchdog alerts, and corruption prevention; a separate anonymisation SOP (v0.8 draft) covers metadata stripping.
How it connects
Scanner and Stain Variability — different scanners produce visibly different images of the same slide, which is the main threat to model generalisation.
De-identification — WSI metadata and label images carry identifiers that survive naive file renaming.
QuPath Annotation Workflow — the tooling layer through which WSIs become annotated research data.
Weakly Supervised Learning — tiling a WSI into instances is the standard way to make these files tractable for machine learning.
Turnaround Time — scanning and transfer time are components of the digital workflow’s overall turnaround.
Mitotic Count — the sharpest example of routine diagnosis depending on WSI metadata: counting per mm² is impossible without the µm/pixel resolution the anonymisation step may be stripping.
Nottingham Grading — reading a slide on a screen changes the diagnosis itself, not only what a model can be trained on; the display becomes part of the measuring instrument.
Video-Based WSI Mosaicing — the same artefact produced without a scanner, by stitching a microscope video. It shows by omission how much of what this page describes is metadata the scanner supplies for free: position, magnification and physical pixel size all have to be reconstructed or supplied by hand.
WSI Format Conversion — what happens when the file is in the wrong format for the tool you need, which is the normal case: the interoperability standard is the one most free viewers cannot open.
Slide Metadata Extraction — the file described here carries its own provenance in its header: scanner, objective power, microns per pixel, scan date. Reading it is a few lines of code, and it is the prerequisite behind most of the open questions in this wiki — without microns per pixel an area-calibrated measurement is not possible at all.
Cervical Cytology Screening — the one application where scanning is a materially different operation. A smear is thick and cells overlap, so a single focal plane does not capture what a screener would see by racking focus; cytology conventionally needs a z-stack, which multiplies scan time and storage per slide and changes every capacity calculation on this page.
Open questions
- Which scanners and at what magnification are used for each of the group’s AI cohorts? The schema asks for this on every model paper and it is not recorded for any of them. [unverified]
- Is DICOM adoption planned, or does the group stay on vendor formats? Relevant to long-term archive strategy. Now priced, though still unanswered. Chauveau 2025 — Converting WSI from DICOM to SVS-like TIFF surveyed the freely available vendor viewers and found only one of six reads DICOM — Aperio ImageScope, NDP.view2 and OlyVia do not, and neither do the named teaching and collaborative-diagnosis platforms. Adopting DICOM therefore means adding a conversion step rather than accepting an unreadable archive. That conversion is lossless and takes seconds per gigabyte, so the cost is real but small; see WSI Format Conversion.
- The
metadata-qupathrepo extracts scanner type and resolution from SVS files, but the ecosystem notes state this metadata is deleted by anonymisation — so is it captured before the anonymisation step? Worth confirming, since it is exactly the field the schema wants.