HistoMIL: A Python package for training multiple instance learning models on histopathology slides.

Area of Science:

• Computational pathology
• Biomedical informatics
• Machine learning in healthcare

Background:

• Hematoxylin and eosin (H&E) stained slides are crucial for disease diagnosis.
• Deep learning advances enable detection of complex molecular patterns in histopathology.
• Multiple instance learning (MIL) shows promise for automated analysis but is complex to implement.

Purpose of the Study:

• To introduce HistoMIL, a Python package streamlining MIL algorithm implementation for computational pathology.
• To facilitate training and inference of MIL-based algorithms for researchers and pathologists.
• To provide an integrated pipeline for feature encoding, transfer learning, and MIL algorithm application.

Main Methods:

• Developed HistoMIL, a Python package utilizing PyTorch Lightning for customization.
• Integrated a self-supervised learning module for feature encoding.
• Implemented a pipeline including transfer learning (TL) and three MIL algorithms: ABMIL, DSMIL, and TransMIL.

Main Results:

• HistoMIL successfully built predictive models for 2,487 cancer hallmark genes using breast cancer histology slides.
• Achieved Area Under the Receiver Operating Characteristic Curve (AUROC) performances up to 85% in gene prediction tasks.
• Demonstrated the package's capability in streamlining complex MIL algorithm workflows.

Conclusions:

• HistoMIL significantly simplifies the application of advanced MIL algorithms in computational pathology.
• The package empowers researchers and pathologists to leverage deep learning for enhanced disease diagnosis and molecular pattern detection.
• HistoMIL facilitates the development of predictive models from histopathology data, advancing precision medicine.