PathFlowAI: A High-Throughput Workflow for Preprocessing, Deep Learning and Interpretation in Digital Pathology.

Area of Science:

• Digital Pathology
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Accurate disease diagnosis relies on analyzing biopsy slides, often requiring precise identification of features within tissue structures.
• Deep learning diagnostic aids show promise for classifying digitized biopsy slides, but require efficient preprocessing workflows for clinical adoption.

Purpose of the Study:

• To present a cost-effective, flexible, scalable, rapid, interpretable, and transparent preprocessing workflow for deep learning diagnostic aids in clinical pathology.
• To demonstrate the utility of this workflow in analyzing liver biopsies for hepatitis evaluation.

Main Methods:

• Developed a preprocessing and deep learning analytics pipeline optimized with Dask and mixed precision training via APEX.
• The workflow handles patch-level or slide-level classification and prediction, incorporating model interpretation and an efficient audit trail.
• Demonstrated the workflow on liver biopsy analysis for hepatitis evaluation from a prospective cohort.

Main Results:

• The PathFlowAI workflow is capable of handling diverse classification and prediction tasks on digitized biopsy slides.
• Preliminary data suggest PathFlowAI is a cost-effective and time-efficient tool for clinical AI applications.
• The workflow demonstrated utility in analyzing liver biopsies for hepatitis.

Conclusions:

• PathFlowAI offers a robust solution for integrating deep learning diagnostic aids into clinical pathology workflows.
• The developed workflow addresses key requirements for clinical adoption, including cost-effectiveness, speed, and interpretability.
• PathFlowAI shows potential to enhance the clinical utility of AI algorithms in anatomic pathology.