About PathoIntern

Phased Approach

A two-phase strategy: validate the core concept first, then extend to full whole-slide image processing.

Phase IActive · 6 Weeks

MVP — Patch-Based Triage

Validate the core concept with pre-extracted patches and public datasets.

Technical Specifications

Input typePre-extracted patches (PNG/JPG)

DatasetKaggle Parasite Dataset

Embedding modelCTransPath — 768-dim

Model accessPublic weights, no approval required

Scoring methodk-NN anomaly detection vs normal baseline

ViewerStatic image display (patch-level)

Data volume< 10,000 patches (MVP scale)

DeploymentDocker Compose (local / dev)

Capabilities

Patch image upload (single or batch)
CTransPath embedding generation and pgvector storage
Criticality scoring with 4-tier classification
Top-k similar patch retrieval with labels
Prioritized worklist dashboard
Pathologist verdict submission (agree/disagree/modify)
Comprehensive audit log with full attribution
Non-dismissible disclaimer on every view

Phase IIPlanned · 8–12 Weeks

Full WSI Integration

Extend to whole-slide images with OpenSlide, deep-zoom viewer, and heatmap overlays.

Technical Specifications

Input typeWhole-Slide Images (NDPI, SVS, TIFF)

DatasetInstitutional library (~150 slides, ~200 GB)

Embedding modelUNI — 4096-dim (HuggingFace)

Model accessRequires HuggingFace access approval

Scoring methodSlide-level aggregated anomaly scoring

ViewerOpenSeadragon deep-zoom viewer + heatmap overlay

Data volumeProduction scale (continuous ingestion)

DeploymentContainerized cloud deployment

Capabilities

Whole-slide image upload and tile extraction
Patch-level UNI embeddings aggregated to slide-level scores
OpenSeadragon deep-zoom interactive viewer
Risk heatmap overlay on whole-slide view
Slide-level worklist with region annotation
Pathologist learning loop (verdict → model update)
LIS/EMR integration API
Multi-institution deployment architecture

Side-by-Side Comparison

Dimension	Phase I (MVP)	Phase II
Timeline	6 weeks (1 sprint/week)	8–12 additional weeks
Input granularity	Patch (cropped region)	Whole slide image
Embedding dimension	768-dim (CTransPath)	4096-dim (UNI)
Primary goal	Validate the core triage concept	Production-grade clinical assistant
Demo quality	High — fully functional end-to-end	Clinical-grade UX
Academic rigor	Moderate (single dataset)	High (multi-dataset + metrics)
Data requirements	Public Kaggle dataset	Institutional WSI library
GPU requirement	Optional (CPU fallback works)	Required for WSI tile processing

Why Two Phases?

Whole-slide image processing requires institutional data access, high-compute infrastructure, and regulatory pathway planning that cannot be completed in a 6-week academic project. Phase I validates the core triage concept — that AI pattern similarity can reliably stratify patches by anomaly risk — with publicly available data and a manageable technical stack.

Phase II then extends the validated concept to clinical-grade deployment, incorporating the feedback collected from Phase I pathologist verdicts and the additional infrastructure required for whole-slide images.

Problem Statement Team