Research tools · Human-centered AI
Multi-Task PathologyOne platform for six annotation workflows.
An annotation platform that turns expert pathology judgments into structured data for AI research.
The challenge: support fundamentally different tasks while keeping decisions focused, progress recoverable, and data consistent.
Problem & role
Turn expert judgments into consistent research data.
The lab connects microscopic patterns with molecular profiles for AI research in cancer diagnosis and prognosis. This depends on consistent labels, validated image regions, and reliable expert records.
The challenge was to unify six annotation workflows without adding cognitive load or losing task-specific context.
Explore the six workflow requirements
The scope
Different judgments.
Different interaction needs.
- 01 / Single-cell morphology
- Rapid evaluation of large cell batches, with decisions saved as work progresses.
- 02 / Feature standardization
- Map heterogeneous labels to a common pathology ontology with suggestions.
- 03 / Spatial ROI triage
- Select relevant reference regions through gated review steps.
- 04 / Ground truth verification
- Visually validate cell and nuclei locations across annotation folders.
- 05 / Cross-cohort validation
- Apply the same schemas to dataset variants for consistency checks.
- 06 / Magnification QA
- Record scanning conditions to preserve context for model training.
Design decisions
Keep the system familiar, while each task sets the pace.
Reveal the next decision
when it is relevant.
Cell evaluation uses sequential gating and per-prompt saving across 36 cells and 13 options. Spatial triage uses Top-K selection with duplicate prevention.
The design balanceGuidance supports focused review, while explicit opt-out paths allow experts to express uncertainty instead of forcing a judgment.
Standardize the label,
not the expert’s judgment.
The mapping workbench brings source labels, standardized terms, and suggestions together. Experts choose the mapping; the shared ontology keeps records consistent.
Make progress
recoverable.
Per-prompt autosave, persistent sessions, and visible completion give experts reliable stopping points during long reviews.
The original platform uses SQLite, thread-safe writes, and CSV snapshots every five minutes to protect annotation records.
Design development
Reuse controls without forcing every task into the same sequence.
A shared input library supports single selection, searchable tags, validated text, range controls, and high-volume lists across six workflows.
The project record documents the implemented system, rather than tested design iterations. The redesign retains shared controls and task-specific progression.
The redesigned workspace below preserves these task patterns. It is a visual and interaction proposal, not evidence of a new validation study.
Original platform implementation
React 18 · TypeScript · Vite
FastAPI · Task-specific REST endpoints · JWT & Bcrypt
SQLite · Automated CSV backups · ML-ready exports
Final experience
Keep evidence and the next decision together.
Switch between annotation tasks without losing the review frame. Original imagery and vocabulary anchor the redesigned controls; selections and saved records remain local to this preview. The navigation groups the source workflows; tabs are not a one-to-one count of the six research requirements.
Open the full annotation workspace ↗Compare regions. Keep the evidence in view.
Calcification present · Review and rank reference regions in sequence.
Define what is not visible.
Review the original 13-feature vocabulary. Selected features are marked as not seen.
Static feature set
6 of 13 not seenDifferent labels. A shared vocabulary.
Review the source label, find a standardized term, and explicitly confirm the mapping.
Inspect the image. Record the context.
Source folder: blood_nuclick · Sample images retain the original annotation overlays.
A clear brief before a careful judgment.
The documented assessment uses 19 questions per bone-marrow cell image.
Focus on the central cell.
The source study describes Wright–Giemsa stained bone-marrow aspirates at 40× magnification.
- Locate the target. Review the central cell, including images with multiple cells.
- Allow an unusable image. If no usable cell is present, answer “No” to the first question.
- Assess the concepts. The 19-point evaluation covers morphology, nuclear features, cytoplasm, granules, and pathology-specific traits.
Ground-truth and cross-cohort review share a schema. Scanning magnification is recorded separately from image display size.
The 19-point bone-marrow assessment retains its documented brief. Unavailable evaluation images and questions are not recreated.
Portfolio demonstration using source screenshot image crops. This preview does not connect to the research platform, produce clinical findings, or reproduce its authentication and storage infrastructure.
Outcome & reflection
A shared workspace, with room for expert judgment.
The platform brings six research workflows into a common system. Shared navigation and inputs provide continuity; task-specific controls preserve the context behind each annotation.
Consistent labels alone do not establish correctness. A next evaluation should examine uncertain cases, disagreement between reviewers, and recovery after interruptions—not just completion.
This is a research annotation tool, not a diagnostic system. No improvement in diagnostic accuracy or reviewer fatigue is established by the visual redesign.
Original platform scope & reported output
The original project reports 50k+ annotated data points. This is a platform-level total, not a count of my personal annotations or a result of this redesign; a supporting statistical breakdown is not available in the project record.
One workflow contains 36 cells and 13 feature options. These describe its configuration, not a measured count of decisions completed in a session.
Next directions
- Multi-task consensus to refine ground truth
- Multi-modal model training
- Real-time quality dashboards
- Expanded solid-tumor support