Slideflow: deep learning for digital histopathology with real-time whole-slide visualization

Area of Science:

• Digital Pathology
• Computational Biology
• Machine Learning

Background:

• Deep learning excels in histopathological image analysis but lacks flexible, open-source tools for interactive deployment.
• Current methods often require specialized environments and extensive data reprocessing, hindering experimentation with new architectures.

Purpose of the Study:

• To develop a flexible, open-source deep learning library for digital pathology, named Slideflow.
• To provide an interactive interface for deploying and experimenting with deep learning models on whole-slide images.

Main Methods:

• Developed Slideflow, a Python package supporting diverse deep learning methods for digital pathology.
• Integrated a fast whole-slide image interface for model deployment and visualization.
• Included tools for optimized whole-slide image processing, stain normalization, augmentation, and weakly-supervised classification.
• Enabled framework-agnostic data processing (TensorFlow/PyTorch) and real-time visualization on various hardware.

Main Results:

• Slideflow facilitates rapid experimentation with new deep learning architectures.
• Achieved highly optimized whole-slide image tile extraction (2.5s/slide at 40x).
• The graphical user interface supports real-time visualization of slides, predictions, heatmaps, and feature spaces.

Conclusions:

• Slideflow offers a flexible and efficient solution for deep learning in digital pathology.
• The library enhances the feasibility and practicality of experimenting with advanced computational methods for histopathological image analysis.
• Its open-source nature and broad hardware compatibility promote wider adoption and collaboration.