HistoMIL

HistoMIL implements Multiple Instance Learning (MIL) pipelines for analyzing Hematoxylin and Eosin (H&E) stained histopathology slides to enable training and inference of models for molecular and phenotypic prediction.

Key Features:

• Self-supervised feature encoding: Incorporates a self-supervised learning module for feature encoding from H&E stained histopathology slides.
• Comprehensive MIL algorithms: Provides Transfer Learning (TL) and three MIL algorithms: Attention-based Multiple Instance Learning (ABMIL), Deep Set-based MIL (DSMIL), and Transformer-based MIL (TransMIL).
• PyTorch Lightning implementation: Implements models and training routines using the PyTorch Lightning framework.

Scientific Applications:

• Computational pathology: Automated analysis of histopathology slides to detect complex molecular and phenotypic patterns.
• Gene-level prediction in breast cancer: Builds predictive models for 2,487 cancer hallmark genes from breast cancer histology slides, achieving Area Under the Receiver Operating Characteristic (AUROC) scores up to 85%.

Methodology:

Implements self-supervised learning for feature extraction, Transfer Learning (TL), and supports ABMIL, DSMIL, and TransMIL implemented on PyTorch Lightning for MIL training and inference.