PathForge

PathForge is a modular benchmarking framework for multiple instance learning (MIL) in computational pathology. It supports whole-slide image (WSI) feature extraction, H5 artifact generation, tile overview reports, MIL benchmarking, pipeline optimization, optional TorchMIL and MIL-Lab backends, metric adapters, and explainability hooks.

Policies and trainers resolve implementations through PathForge interfaces and registries, while concrete third-party integrations are concentrated in adapter modules where practical.

---

Getting Started

• Introduction
• Installation
• Data Preparation
• Quick Start

Tutorials

• Tutorials
  • End-to-End Classification Tutorial
  • Tutorial: Feature Extraction
  • Tutorial: Benchmarking
  • Tutorial: Pipeline Optimization
  • Tutorial: Inference and Heatmaps
  • Tutorial: Model Packaging After Benchmarking / Optimization
  • Tutorial: Slide Retrieval
  • Tutorial: Command-Line Interface
  • Workflow Overview

Reference

• Configuration Reference
• MIL Benchmark and Optimization Options
• Testing
• Backends
• Architecture
• PathForge HDF5 Structure — Canonical Overview
• Slide retrieval
• PathForge Retrieval Representation H5 Structure
• Slide Retrieval Results And Metrics
• Task Outputs, Metrics, and Visualizations
• Supported slide files and slides_dir layout
• Scaling and Cluster Execution
• Troubleshooting

API Reference

• API Reference

---

Key Capabilities

Feature	Description
Feature extraction	Tile WSIs, segment tissue, extract tile features, persist row-aligned H5 artifacts.
Benchmarking	Grid-search over model, loss, feature extractor, activation, and optimizer combinations.
Optimization	Optuna-driven pipeline search with configurable samplers and pruners.
Inference	Checkpoint-based prediction and per-instance attention heatmap generation.
Backends	Native PathForge models, TorchMIL, or MIL-Lab via backend adapters.
Metrics/losses	Optional TorchMetrics (classification) and TorchSurv (survival) integrations.
Explainability	Per-instance MIL attention heatmaps stored alongside slide H5 artifacts.
TCGA integration	Direct dataset download via tcga-tools, auto-generated annotation CSVs.