AI Pathology Diagnosis System (DECIMA)
Computer Vision & LLM Clinical Decision Support
Automated hematology cell counting and disease-pattern classification reducing review time from 5 hours to 2 minutes.
The Engineering Challenge
Clinical diagnostic laboratories faced heavy backlogs and diagnostic fatigue reviewing hundred-slide peripheral blood smears.
Core ProblemManual microscope counting of WBCs, RBCs, and platelets was tedious, subjective, and prone to intra-observer variability.
- High accuracy across staining variations and slide preparation artifacts
- Strict positioning as decision-support software (never unattended clinical diagnosis)
- Structured LLM report generation aligned with laboratory reporting standards
System Architecture
End-to-end data pipelines and component isolation boundaries.
High-resolution image tiling pipeline feeding deep computer vision cell classification models, combined with an LLM clinical report generator synthesizing morphology findings into standard hematology summaries.
Key Engineering Decisions
Two-Stage Detect-Then-Classify Architecture
Separated cell boundary localization from fine-grained morphological classification to handle crowded cell clusters.
Strict Decision-Support Grounding
Enforced deterministic schemas on LLM report outputs with confidence intervals to ensure complete transparency for reviewing pathologists.
Production Verification & Results
- Achieved 90% classification accuracy benchmarked against 100+ pathologist-labelled reference slides
- Reduced pathologist review turnaround time from 5 hours to 2 minutes per case
- Generated structured clinical reports with automated cell counts and morphology warnings
Production Stack & Tooling
Next Iterations & Roadmap
- Multi-modal fusion incorporating patient EHR history with digital slide images
- Active learning pipeline flagging ambiguous cells for specialist review
Want to build something similar?
I specialize in taking complex agentic workflows, duplex voice architectures, and high-throughput computer vision models into production.