Skip to content
All work

AIGermany2024

Tumor-cell detection for digital slides pathology

Gigapixel whole-slide segmentation and cell scoring — reproducible and audit-ready.

Client
Mindpeak
Region
Germany
Sector
Computational pathology / vision
Engagement
2024 · 23 weeks
Team
3 ML · 1 MLOps
Status
In validation

The challenge

Pathologists scoring biomarkers on whole-slide images work at gigapixel scale by hand, which is slow and varies between readers. Mindpeak needed consistent tumor-cell detection and scoring that runs at slide scale under a regulated, reproducible pipeline.

What we built

The full pipeline, end to end — not a blurb.

  1. 01Vision

    Tile-based segmentation and cell-classification models with stain normalization.

  2. 02Scale

    A whole-slide inference pipeline with tiling, stitching, and GPU batching.

  3. 03MLOps

    Dataset versioning, reproducible training, and a model audit trail aligned to IEC 62304.

  4. 04Deployment

    Deterministic engines with a reviewer overlay and scores.

Results

38 s

gigapixel whole-slide inference, from 6+ minutes of manual regions

0.93

cell-detection F1 on the held-out cohort

46%

reduction in inter-reader scoring variance

3

scanner vendors validated against

In clinical validation with partner labs.

Have a constraint like Mindpeak’s? Bring us yours.

Start a project

Next engagement

Blues Wireless

Cellular IoT device + fleet