Multi-Task Pathology
Annotation Platform
A unified interface streamlining 6 distinct workflows, from cell morphology to spatial triage, into one intuitive, cognitively optimized system.
Research Motivation
Fueling AI-Driven Pathology
To support the lab's mission of establishing robust and fair AI methods, this platform was engineered to generate the high-quality, quantitative ground truth required to train fully automated algorithms for cancer diagnosis and prognosis.
Bridging Pathology & Multi-Omics
The lab aims to connect molecular profiles with microscopic patterns. My platform enables this by standardizing morphological labels, providing the semantic structure needed to elucidate the molecular aberrations underpinning cancer cell diversity.
Enabling Real-World Clinical Integration
Developing multi-modal AI models for diverse populations requires rigorous data validation. This system unifies 6 distinct workflows to ensure that the integrated pathology and clinical data is precise enough to inform treatment decisions.
The Challenge
Design a single annotation platform that unifies six fundamentally different pathology workflows—from high-speed cell evaluation to strategic spatial triage—while strictly controlling cognitive load to ensure consistent, high-quality training data.
Six Distinct Annotation Workflows
Single-Cell Morphology
High-frequency evaluation of large cell batches with rapid, auto-saved decisions.
Clinical Ground Truth Verification
Visually validating cell and nuclei locations across annotation folders.
Pathology Feature Standardization
Mapping heterogeneous labels to a standardized pathology ontology with smart suggestions.
Cross-Cohort Validation
Reapplying identical schemas to dataset variants for consistency checks.
Spatial ROI Triage
Strategically selecting the most relevant reference regions through gated review steps.
Magnification QA
Logging scanning conditions to ensure data integrity for model training.
Core Design Constraints
Cognitive Load Reduction
Prevents decision fatigue via gated steps.
Standardized interaction patterns across tasks.
Smart suggestions to reduce manual input.
Task-Specific UI Patterns
Auto-saving workflows for high-volume tasks.
Progressive disclosure for complex decision trees.
Dynamic triage panels for spatial annotation.
Unified Data Architecture
Scalable architecture supporting 6 distinct workflows.
Robust state management (Session Persistence).
Single source of truth for all annotation types.
ML-Ready Data Infrastructure
Structured exports for immediate model training.
Consistent ontology mapping across all tasks.
Reproducible datasets via versioned APIs.
Design Goal
Create a unified platform that minimizes cognitive load across six distinct workflows and guarantees data quality via structured exports—enabling efficient evaluation while maintaining rigorous diagnostic accuracy.
The Solution
A unified platform that adapts its interaction patterns to match the specific cognitive demands of six distinct diagnostic workflows, ensuring efficiency and data consistency across varying contexts.
High-Frequency Exam Interface
-
36 cells × 13 options with sequential gating to prevent fatigue.
​​
-
Auto-save per prompt to allow resumption from any cell.
Abstract Mapping Workbench
-
Smart autocomplete suggestions for standardized ontology mapping.
Strategic Triage & Review
-
Top-K selection with duplicate-prevention for spatial reference tiles.
​​
-
Tiled grid layout for fast visual review of image tiles.
Modular Input System
A shared library of standardized components that powers all six workflows. This unified design system ensures consistent interaction patterns across diverse tasks, significantly reducing the pathologist's learning curve.
Single Selection
Radio & Dropdown
Searchable Tags
Multi-Select Combobox
Inline Text
Input Fields & Validation
Range Selector
Slider & Stepper
High-Volume List
Virtual Scroll / Pagination
Zero-Loss Data Persistence
A fault-tolerant system that captures every input in real-time. Pathologists can pause complex workflows, switch tasks, or disconnect instantly with full session history preserved—ensuring total peace of mind during long evaluations.
✅ Seamless Session Resumption
✅ Real-Time Synchronization
✅ Automated Redundant Backups
System Architecture
Modular React frontend • FastAPI with task-specific REST endpoints • SQLite persistence • Automated CSV backups
React 18
TypeScript + Vite
FastAPI
Python Backend
SQLite
Local Database
Authentication
JWT & Bcrypt
CSV Export
ML Pipeline
Multi-Task Data Pipeline
Morphology
Label Mapping
Spatial Triage
Unified Application Layer
React Frontend
FastAPI Validation
Persistence & Export
SQLite DB
ML-Ready CSV
Robust Data Persistence
✅ A relational design that normalizes shared metadata (user, session, time) while accommodating flexible payload structures.
✅ Thread-safe writes ensure no data is lost even during rapid-fire annotation sessions.
✅ Background processes generate CSV snapshots every 5 minutes for redundancy.
Impact & Outcomes
A unified platform generating structured, multi-modal pathology annotations across six distinct clinical workflows
6
Unified Workflows
Consolidated fragmented tools—from micro-level single-cell exams to macro-level spatial triage—into one cohesive, streamlined ecosystem.
50k+
Annotated Data Points
Generated a massive repository of structured, multi-modal annotations, providing the critical ground truth needed to train robust pathology AI models.
~500
Decisions per Session
Optimized the interface to handle high-volume data capture (36 cells × 13 options) without inducing user fatigue or compromising accuracy.
Cross-Workflow Design Principles
Context-Aware Consistency
Tailored workflows share a common design language. Task-specific interfaces are built on a modular component system, ensuring familiarity across different annotation modes.
Cognitive Safety & Disclosure
Complex decisions are revealed step by step. Progressive disclosure reduces overload and provides clear opt-out paths when uncertainty arises.
Fault-Tolerant Continuity
All actions are saved instantly. Users can pause, switch tasks, or disconnect at any time without risking data loss.
Unified Data Normalization
A single architecture supports diverse data types, normalizing all outputs into a consistent, ML-ready format.
Future Roadmap
Refining Ground Truth
Multi-Task Consensus
Closing the AI Loop
Multi-Modal Training
Real-Time Insights
Quality Dashboards
Domain Expansion
Solid Tumor Support
Final Outcome
Login Page
Track real-time progress across all workflows. View completion rates at a glance and resume tasks exactly where you left off.
Task 1: Static Options
Transform chaotic labels into structured data. Use smart suggestions to map diverse terms to a unified ontology with zero typing errors.
Task 3: Image Annotation
Ensure training granularity. Explicitly classify targets as 'Cell' or 'Nuclei' and verify masks to create a gold-standard dataset.
Single Cell Patch Concept Matching
A secure, JWT-protected gateway that instantly restores your workspace. You can pick up exactly where you left off.
Personalized Task Dashboard
Navigate diagnostic queues instantly via the sidebar. Auto-save lets you switch between hundreds of cases seamlessly, with no manual saving required.
Task 2: Label Mapping
Curate high-value data from massive images. The grid view enables rapid 'Top-K' selection, preserving only the most clinically relevant regions.
Task 4-1: Cell Image Annotation
Validate model predictions against expert judgment. A rigorous 19-point assessment confirms AI-generated concepts on bone-marrow cells.