Back to portfolio

Case study · Pathology Lab · Apr 2025 – present

SPAI: AI Cancer Diagnostics Tool (Digital Pathology)

A pathology AI web tool for colorectal polyp whole-slide image analysis.

Role
AI / ML Engineer
Building
Research Laboratory
Year
Apr 2025 – present

Problem

Pathologists reviewing colorectal polyp whole-slide images need faster, consistent support for atypical vs benign and tubular adenoma vs hyperplastic classification — without replacing clinical judgment.

Daniel's role

Deployed and integrated the lab's PyTorch models on AWS SageMaker and EC2. Fixed GPU out-of-memory crashes with tile streaming, and wired Fast (128 tiles) vs Full (whole-slide) analysis modes into the product workflow.

System diagram

SPAI digital pathology pipelineWSISageMakerTile streamModesFast / Full

Stack

  • Python
  • PyTorch
  • AWS SageMaker
  • EC2
  • Tile streaming
  • Whole-slide imaging

Outcome

Shipped Fast and Full analysis modes. Daniel organized a pathologist error-case review. Separately, the lab's model accuracy rose from 76% to 88%.

Clinical constraints shape deployment work — latency, memory limits, and workflow fit matter as much as raw model scores.

Field notes (measurements)

  • Modes

    Fast / Full

  • Lab accuracy

    76% → 88%

Find this building on the city plan →