Computational pathology research system

See what the model sees.

From pathology slide to verified evidence — an inspectable multi-agent system for computational pathology.

Launch interactive demo
Case
GBX-BR-001
Organ
Breast
Status
Ready for investigation
Scroll to enter the glassbox
LEVEL 3 · 2.5×
LEVEL 2 · 5×
LEVEL 1 · 10×
WHOLE SLIDE IMAGE · LEVEL 0 · 40×
98,304 × 62,464 PX · DEMO
0.25 µM/PX
GBX-BR-001 · H&E
DEMO DATA
TILE T-0642 · X 41216 · Y 28160 · 512²
SAM2 · MASK LAYER +0.4U
SPATIAL EXPRESSION LAYER · DEMO
REGION
R-016
Mask confidence0.94Area4 210 µm²Coordinates41 216, 28 160SourceSAM2 · ft-path
CHAPTER 01
The slide enters the Glassbox.

One demonstration case. Every transformation that follows happens to this same object — nothing is swapped out behind the scenes.

CaseGBX-BR-001OrganBreastStainH&EStatusReady for investigation
AGENT 01 · SLIDE INTELLIGENCE
Preparing the slide for analysis.

A whole slide image is not one picture. It is a resolution pyramid — the same tissue held at several magnifications, indexed by coordinate.

WSI loadedOKPyramid levels0Tissue detected61.4%
AGENT 01 · TILING
One slide becomes addressable regions.

Tiles separate along their true coordinates. Patches without tissue fade out — they are never sent downstream.

Tiles generated0Tissue retained0Coordinates indexedOK
Mapping slide coordinates…
AGENT 02 · SAM2 SEGMENTATION
Segment what matters.

A pathology-adapted, fine-tuned SAM2 model turns visual tissue into spatially addressable evidence. Masks sit as thin translucent layers a fraction of a unit above the tissue.

Regions segmented0Mean mask confidence0.91 demo
AGENT 03 · TRANSCRIPTOMICS
Connect morphology with molecular evidence.

Expression measurements enter as a third plane in the same coordinate system. Different forms of evidence, one biological region.

Spots aligned40 demoSignature— placeholderSpatial agreement— placeholder
THE GLASSBOX MOMENT
Every layer that produced the answer, still visible.
AGENT 04 · VERIFICATION
AWAITING EVIDENCE
Question the conclusion.

Before evidence becomes a report, it must survive verification. Confidence is reported as components, never as one number.

Model confidence—
Evidence agreement—
Spatial consistency—
Image quality—
Transcriptomic agreement—
OOD risk—
Calibrated confidence—
Awaiting evidence bundle…
ADDITIONAL INVESTIGATION · VERIFIER RETURNS TO THE EVIDENCE PIPELINE
A04→A01Request region R-041
A01Tile T-0911 loaded
A02SAM2 re-segmentation
A03Expression re-associated
AGENT 05 · REPORT GENERATION
DEMO DATA
Evidence becomes explanation.
CASE SUMMARY

Demonstration case GBX-BR-001, breast tissue, H&E. Analysis performed on tile T-0642 with one additional region requested during verification.

MICROSCOPIC FINDINGS

Placeholder finding text. Morphological description will be supplied by the pathology content model.

MORPHOLOGICAL EVIDENCE

12 regions segmented; mask confidence and area recorded per region. Values shown are demonstration placeholders.

TRANSCRIPTOMIC EVIDENCE

Expression layer aligned to the same coordinate space. Signature association pending scientific specification.

VERIFICATION & CONFIDENCE

Initial bundle failed verification on evidence agreement; re-investigation of R-041 resolved the disagreement. Reported as components, not a single score.

EVIDENCE TRACE
Select a reference in the report to travel from the conclusion back to the pixels.
SLIDE
Whole slide image
GBX-BR-001
TILE
Selected tile
T-0642
SAM2
Segmentation mask
R-025
TRANSCRIPTOMICS
Expression layer
40 spots
VERIFICATION
Verification state
verified · demo
Selected reference
E-04
From pixels to evidence.
From evidence to trust.

Glassbox-Pathology makes the computational pathology pipeline inspectable from the original slide to the final report.

Launch interactive demoExplore architecture
Agentic System · Section 4

Five agents exchanging structured evidence.

Evidence flows forward from WSI tessellation to report synthesis. When verification challenges fail, the orchestrator loops back to collect targeted regional evidence before issuing a verdict.

Explore Models →
Active Mode: Forward Multimodal Evidence Pipeline
Click any agent card for deep architectural specifications
AGENT 01
Orchestrator
Slide Intelligence & ReAct Loop

Indexes gigapixel whole-slide pyramids, extracts candidate tissue patches, and coordinates the multi-agent investigation state machine.

OpenSlideOtsu FilterReAct Controller
in WSI → out Tile[]INSPECT
AGENT 02
Visual Analysis
Morphology & SAM2 Segmentation

Extracts deep visual embeddings and segments individual cell nuclei and tissue boundaries into addressable masks.

UNI BackboneSAM2 (HEST fine-tuned)CellViT
in Tile → out Mask[]INSPECT
AGENT 03
Evidence Investigation
Transcriptomics & Multimodal Fusion

Aligns spatial transcriptomic barcode sequencing with morphology to surface diagnostically relevant regions.

CLAM Attention MILk-NN RetrievalVisium Spatial
in Region → out Evidence[]INSPECT
AGENT 04
Diagnosis and Verification
Adversarial Calibration & Loopback Gate

Formulates diagnostic hypotheses and adversarially challenges them across 7 confidence components, triggering loopback on discordance.

Conformal Predictor (90%)Temperature ScalingMahalanobis OOD
in Bundle → out VerdictACTIVE
AGENT 05
Explanation and Report
Evidence-Grounded Reporting & Q&A

Synthesizes formal 6-section clinical diagnostic reports citing verified evidence and powers interactive clinician Q&A.

CONCH BioGPT TokenizerEvidence-Constrained RAG
in Verified → out ReportINSPECT
Verification Loopback PipelineAgent 04 → Agent 01
When Agent 04 detects low evidence agreement or out-of-distribution (OOD) risk, it rejects the provisional diagnosis and commands Agent 01 to load and inspect adjacent contiguous or margin regions.
AGENT 04 · ARCHITECTURAL SPECIFICATION
Diagnosis and Verification(Adversarial Calibration & Loopback Gate)

Generates hypotheses via organ-specific heads on frozen UNI embeddings and independently challenges them. Evaluates 7 metrics: model confidence, evidence agreement, spatial consistency, image quality, transcriptomic concordance, OOD risk (Mahalanobis distance), and calibrated confidence (conformal prediction with 90% coverage). If discordance arises, it rejects the diagnosis and dispatches a re-investigation loop to Agent 01.

Input Data Contract
Multimodal Evidence Bundle
Output Payload
Calibrated Verdict / Loopback Request
Model Architectures
Conformal Predictor (90%)Temperature ScalingMahalanobis OOD
PIPELINE

Every block is inspectable.

Select a stage to view its architectural purpose, data contracts, and current implementation status.

STAGE 03 · agent 02
SAM2 segmentation

Fine-tuned SAM2 produces promptable contours for cell nuclei and glandular units.

InputTile patchOutputSegmentationMask[]ModelSAM2 (fine-tuned on HEST)StatusImplemented
RESEARCH PRINCIPLES

Why a black-box diagnosis is not enough.

Proposal & Delta vs SlideSeek →Datasets & HEST-1k
01 · PROBLEM
A label is not an explanation

A model that outputs a class gives a clinician nothing to interrogate. The reasoning path, not the prediction, is what can be reviewed.

02 · APPROACH
Make the pipeline the interface

Every intermediate artefact — tile, mask, expression measurement, verification score — is a first-class object the user can open.

03 · MULTIMODAL
One coordinate system

Morphology, segmentation and transcriptomics describe the same region. Holding them in a shared space is what makes agreement measurable.

04 · VERIFICATION
A conclusion must survive challenge

The verifier is adversarial by design: it looks for disagreement and can send the pipeline back for more evidence.

05 · CALIBRATION
Confidence as components

Model confidence, evidence agreement, spatial consistency, quality and OOD risk are reported separately. Implementation to be supplied.

06 · LIMITATIONS
Nothing here is a result

This site is a design and architecture prototype. All values are demonstration placeholders pending the scientific implementation.