ChampKit: A framework for rapid evaluation of deep neural networks for patch-based histopathology classification

Area of Science:

• Digital pathology
• Computer vision
• Machine learning

Background:

• Histopathology is crucial for cancer diagnosis.
• Deep learning advances image analysis but optimal model selection is challenging.
• Lack of systematic evaluations hinders progress in histopathology classification.

Purpose of the Study:

• Introduce ChampKit, a user-friendly software toolkit for evaluating neural network models in histopathology.
• Enable robust and systematic assessment of deep learning models for patch classification.
• Facilitate model selection for algorithm developers and biomedical researchers.

Main Methods:

• ChampKit offers an extensible, reproducible toolkit for training and evaluating deep neural networks.
• It supports a wide range of public datasets and models via command-line interface.
• Enables integration of external models with minimal coding.

Main Results:

• Systematic evaluation of multiple neural networks across six datasets was performed using ChampKit.
• Transfer learning showed benefits primarily in low-data regimes.
• Self-supervised pretraining unexpectedly yielded limited performance improvements compared to other methods.

Conclusions:

• Selecting appropriate models for digital pathology datasets is complex.
• ChampKit democratizes the evaluation of deep learning models for various pathology tasks.
• The toolkit and its code are publicly available for broader scientific use.