PathFlowAI

PathFlowAI processes digitized whole-slide images and associated annotations to enable high-throughput preprocessing, deep learning-based patch- and slide-level classification, and interpretable visualization for pathology analysis.

Key Features:

• High-Throughput Workflow: Capable of processing more than 500 whole-slide images per day for large-scale histopathology datasets.
• Parallel Preprocessing with Dask: Uses Dask to preprocess annotations and whole slide images in parallel.
• Deep Learning Model Training: Trains deep learning models for patch-level and slide-level classification using mixed precision training via APEX.
• Interpretability with SHAP: Applies SHapley Additive exPlanations (SHAP) to identify and highlight regions that contribute to model predictions.
• Visualization with UMAP embeddings: Employs Uniform Manifold Approximation and Projection (UMAP) embeddings to visualize complex feature and sample relationships.
• Storage-Efficient Audit Trail: Records processing steps in a storage-efficient audit trail to document preprocessing and analysis operations.

Scientific Applications:

• Anatomic pathology specimen analysis: Applied to analysis of anatomic pathology specimens, including liver biopsies for evaluation of hepatitis.
• Prospective cohort evaluation: Demonstrated utility on a prospective cohort study to assess performance on clinical-style datasets.

Methodology:

Preprocessing of whole slide images and annotations with Dask; training of patch- and slide-level deep learning models using mixed precision via APEX; interpretation with SHAP and visualization with UMAP; and maintenance of a storage-efficient audit trail.