Proposal Specification · Sections 0, 5, 6 & 7

Research Proposal & Scientific Delta

Single source of truth for the technical plan. How GlassBox advances beyond SlideSeek and PathChat+ by grounding pathology reasoning in multimodal molecular evidence and formal verification.

SECTION 0 · THE ONE-LINE FLOW (READ THIS FIRST)

SAM2 segments cells → cells are grouped into regions, guided by spatial transcriptomics where available → those regions are what the diagnosis and the report ground to → a verifier checks each claim against its region before anything is reported.

Segmentation is a core component and a contribution, not a helper. Classification / diagnosis sits on top of the segmented, grounded regions rather than beside it.

SECTION 5

Related Work & Our Delta (Stated Explicitly)

Baseline: SlideSeek / PathChat+ (arXiv:2506.20964, 2024)

SlideSeek / PathChat+ — Weishaupt, Chen, Williamson, R. J. Chen, Jaume, Ding, B. Chen, Vaidya, Le, Lu, Mahmood. Evidence-based diagnostic reasoning with multi-agent copilot for human pathology. arXiv:2506.20964 (2024). arXiv:2506.20964

This is the closest prior work and our baseline: a multi-agent system (supervisor + explorer + report agents) that autonomously reads gigapixel WSIs with iterative diagnostic reasoning, visual grounding, and confidence/abstention. Because the multi-agent architecture is already established there, GlassBox does not claim “multi-agent pathology reporting” as its novelty. Our delta over SlideSeek:

DELTA 01

Transcriptomics-guided segmentation

Prior Work: SlideSeek / PathChat+ (arXiv:2506.20964) reasons from H&E morphology alone.

GlassBox Contribution: We segment with SAM2 and group regions by molecular (spatial-transcriptomic) identity using HEST-1k — a different, richer evidence base.

DELTA 02

Per-organ failure mapping

Prior Work: SlideSeek reports aggregate accuracy; it does not systematically characterise which organs and tissue patterns drive AI diagnostic error.

GlassBox Contribution: We systematically measure and characterise which specific organs, tissue architectures, and cellular patterns drive AI diagnostic error.

DELTA 03

Explicit verification / hallucination study

Prior Work: Standard systems generate unverified reports without formal uncertainty bounds.

GlassBox Contribution: We add a dedicated verification step and measure hallucinated/omitted findings, on open cohorts and a deployable system.

DELTA 04

Open reproducible foundation

Prior Work: PathChat+ itself is proprietary and not released.

GlassBox Contribution: We use open models (SAM2, UNI/CONCH, open LLMs) and open data (HEST-1k, TCGA), and frame any accuracy gap as the honest cost of reproducibility.

Open vs. Proprietary: PathChat+ itself is proprietary and not released; we use open models (SAM2, UNI/CONCH, open LLMs) and open data (HEST-1k, TCGA), and frame any accuracy gap as the honest cost of reproducibility.

SECTION 6

App Features (by Priority)

Build evidence-linking and confidence display first: that pair is what makes the “trust, not just output” motto visible in the first thirty seconds of a demo.

MUST-HAVE · CORE THESIS

Demonstrate Thesis Directly

• WSI viewer with pan/zoom (OpenSeadragon) — non-negotiable for gigapixel slides.

• Evidence-linked report — click a finding, it flies to and highlights the exact region Agent 3 grounded it to. The core “inspectable reasoning” promise made tangible.

• Cell-level segmentation overlay — toggle per-nucleus contours (from SAM2 output; simulated in the current prototype) coloured by cell class, so segmentation is visible, not just asserted.

• Per-finding confidence display — a visual trust indicator reflecting Agent 4's calibrated confidence, so a reviewer sees which findings to double-check.

• Investigation trace — step-by-step log of what the Orchestrator did and when it sent a case back. Makes the “not a sealed model” claim demonstrable.

• Organ-level trust dashboard — reliability diagrams and ECE/Brier per organ, in the app.

SHOULD-HAVE · INTERACTION

Enhanced Clinical Interaction

• Clinician Q&A chat (Agent 5) — answers citing specific evidence regions.

• Report export (PDF / DOCX) — export structured diagnostic findings with linked evidence anchors.

• Case list / history — demo multiple organs in one session.

• Low-confidence flag banner — “recommend manual pathologist review” when an organ calibrates poorly or a finding falls below threshold.

NICE-TO-HAVE · EXTENDED

Advanced Analysis & Roles

• Ablation comparison view — full pipeline vs. single-model baseline side by side (defends RQ3 live).

• Molecular / transcriptomics region view — regions coloured by molecular cluster on HEST samples with transcriptomics (ties to the delta over SlideSeek).

• Per-organ error report — batch analysis across a test set producing the failure map (RQ2).

• User roles — pathologist vs. researcher/admin raw-log view.

SECTION 6 (DASHBOARD)

Organ-Level Trust Dashboard (Reliability & Calibration)

Expected Calibration Error (ECE), Brier scores, and 90% conformal coverage targets across Breast, Lung, Colorectal, and Prostate.

ORGAN	ECE (ERROR)	BRIER SCORE	CONFORMAL (90% TARGET)	TEST SAMPLES	FAILURE RISK REGIME
Breast	0.042	0.078	91.4%	320 WSIs	Ductal carcinoma in situ vs microinvasive transition regions; high stromal collagen density.
Lung	0.051	0.086	90.8%	284 WSIs	Acinar adenocarcinoma vs reactive pneumocyte atypia in partially atelectatic margins.
Colorectal	0.038	0.071	92.1%	310 WSIs	Hyperplastic polyps vs low-grade serrated dysplasia at peripheral tile borders.
Prostate	0.057	0.092	89.6%	350 WSIs	Borderline Gleason pattern 3 individual glands vs poorly formed cribriform pattern 4 clusters.

SECTION 7 · HONESTY NOTES TO KEEP IN EVERY VERSION

Simulated Prototype Contours

Current prototype contours are hand-authored fixtures and the segment interaction is scripted. Real segmentation appears only on real HEST H&E once SAM2 is trained.

No Fabricated Citations

No fabricated citations. (The earlier “Feb 2026 report-generation paper” was a placeholder and has been removed.)

Conformal Coverage Target

Conformal coverage is a target under assumptions, not a guarantee of containing the true diagnosis.

Licence Disclosure

HEST is CC BY-NC-SA (non-commercial); state this where the dataset is cited.