COMPUTATIONAL_PATHOLOGY
DigitalPathology&Oncology
WSI Analysis at Lab Scale
Whole-slide image analysis with QuPath v0.6.0. Tumor detection, TIL scoring, PD-L1 quantification with batch pipelines and deep-learning integration.
8.4
GB/slide
124,893
tiles
34.7
% burden
QuPath • Berkeley • Meta AI
"
Pathology labs must turn whole-slide images (1–20 GB per slide) into actionable, quantifiable findings at scale — tumor burden, margins, TIL density, PD-L1 scoring, grading
LEGACY_WORKFLOW_ISSUES
01
Slow & Manual
Non-reproducible workflows
02
Hard to Audit
No governance or versioning
MODEL_DEPLOYMENT_GAPS
03
Can't Ingest WSIs
Prototypes stall on I/O at scale
04
No LIMS Integration
Can't plug into lab workflows
Deterministic tile I/O → model inference → QC → report on multi-TB WSI cohorts with versioned scripts, audit logs, and batch governance
QuPath
Version 0.6.0 (2025)
Open-source bioimage analysis for research
Cross-platform WSI analysis workstation
Interactive ML & deep-learning hooks (DJL)
Scripting & batch pipeline automation
Cell detection & pixel classifiers
TMA dearraying & spatial analyses
Exportable tables for LIMS integration
PRIMARY_REPO
github.com/qupath/qupath
Core WSI analysis platform with scripting and ML integration
OPTIONAL_ML_BACKEND
github.com/TissueImageAnalytics/tiatoolbox
PyTorch WSI tooling for tiling, stain normalization, inference
Production Workflows
QuPath v0.6.x
[A]
Tumor Detection & Burden
H&E; Solid Tumors
A
PIPELINE
Tissue mask → tiling → model inference → probability heatmap → tumor area % per slide
QUPATH_FEATURES
Pixel classifier, DL via DJL, overlay rendering, batch export
[B]
TIL Density & Spatial Metrics
Tumor-Infiltrating Lymphocytes
B
PIPELINE
Cell detection/classification → Voronoi/Delaunay spatial stats → TIL density per mm²
QUPATH_FEATURES
Object classifiers, cell segmentation, spatial analyses, batch tables
[C]
PD-L1 Scoring (IHC)
Immunohistochemistry
C
PIPELINE
Stain deconvolution → membrane positivity per tumor cell → CPS/TPS computation
QUPATH_FEATURES
Stain vectors, intensity thresholds, object measurements & scripts
[D]
TMA High-Throughput Scoring
Tissue Microarray
D
PIPELINE
TMA dearraying → per-core detection/quantification → export
QUPATH_FEATURES
Built-in TMA dearraying and batch processing
[E]
Serial Section Registration
H&E ↔ IHC
E
PIPELINE
Section registration → transfer ROIs → quantify markers in matched regions
QUPATH_FEATURES
External registration (TIAToolbox) + QuPath ROI operations
Technical Stack
Data • I/O • Performance
WSI Formats
Supports major vendors; tile-wise streaming with caching
Batching
Headless QuPath runs via CLI + script files for thousands of slides
Outputs
Overlay masks, detections, CSV tables, screenshots, project summaries
DEEP_LEARNING_INTEGRATION
In-QuPath
DJL backends (PyTorch/ONNX)
Model pack + script for tile scoring
DJL engines with model pack + script
External
TIAToolbox
Export tiles → inference → re-import masks/GeoJSON
QuPath exports → TIAToolbox inference → re-import
V
Versioned Scripts
QuPath project + git for SOPs + deterministic seeds
A
Audit Trail
Parameters + model hash in exported tables and logs
Q
QA Process
Blinded double-read, discordance dashboards, acceptance gates
Case Blueprints
1
Solid Tumor Burden at Cohort Scale
GOAL
Quantify tumor % across 2,000 WSIs for trial stratification
BUILD
QuPath project → tissue mask (pixel classifier) → model via DJL → overlay + cleanup → batch export
SLOs
≥150 slides/hour/node headless • Re-run deterministically with fixed seeds
ARTIFACTS
CSV per slide/ROI • Overlays for QC • Run logs
2
PD-L1 IHC Scoring with Reviewer QA
GOAL
Standardized TPS/CPS with audit trail
BUILD
Stain deconv + membrane scoring + tumor cell gating • Reviewer verifies in QuPath • Discrepancies flagged
SLOs
TPS/CPS ICC ≥0.85 vs consensus • <5% rescans • Batch run nightly
ARTIFACTS
CSV to LIMS • QC overlays • Reviewer concordance report
3
TIL Density + Spatial Heterogeneity
GOAL
Derive TIL density and tumor-border proximity features for survival modeling
BUILD
QuPath cell detection + classifier • Polygonized tumor regions • Spatial stats • Export features
SLOs
Feature table ready T+24h • QC overlays retained in project
ARTIFACTS
Feature table to R/Python • Spatial heatmaps • Cell counts
Validation & Operations
VALIDATION_METRICS
Tumor segmentation
Dice / Jaccard, pixel AUC
Heatmap calibration & threshold sweeps
Cell detection
F1 / mAP
Class-wise counts, error analysis per stain
PD-L1
TPS/CPS agreement, ICC
ROI-specific audits; reviewer consensus
TIL
Density/mm², spatial proximity
Robustness to staining/section drift
Throughput
Slides/hr, P95 latency
Headless batch with tile cache
Robustness
Stain drift tests
Cross-site WSIs, scanner models
OPS_INTEGRATION
1
Storage
QuPath projects under git/LFS; WSIs on NAS/object storage
2
Scale-Out
Split project per cohort; headless jobs per node; consolidate tables
3
Security
Workstation/offline or VPC; no PHI leaves lab unless de-identified
4
LIMS Handoff
CSV/JSON exports with case IDs; Python ETL to trial DBs
HEADLESS_BATCH_EXAMPLE
QuPath-0.6.0/binaries/QuPath \ --script run_pipeline.groovy \ --project my_project.qpproj # Use Groovy/QuPath scripts to: # - Open project # - Iterate entries # - Run detectors/classifiers # - Save overlays # - Export tables
From whole-slide images to quantifiable findings
SLIDES/HR
≥150
DICE_SCORE
≥0.85
ICC
≥0.85
LATENCY_P95
<60s
✓
QuPath v0.6.0 (2025) — actively maintained
✓
Deep-learning integration via DJL
✓
Batch pipelines for lab-scale processing
✓
Optional TIAToolbox for PyTorch models
QuPath • TIAToolbox • Scripting • Batch Processing • LIMS Integration