Multi-scale representation attention based deep multiple instance learning for gigapixel whole slide image analysis.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Convolutional neural networks (CNNs) are increasingly used for tumor diagnosis using whole slide images (WSIs).
• Training CNNs often requires slide-level labels, but handling gigapixel WSIs presents challenges due to their size and intra-image variations.
• Existing methods struggle with the direct analysis of large-scale WSIs.

Purpose of the Study:

• To propose a novel, end-to-end interpretable deep multiple instance learning (MIL) framework for whole slide image (WSI) analysis.
• To overcome the challenges associated with directly handling gigapixel WSIs in tumor diagnosis.
• To enhance classification accuracy and model interpretability in WSI analysis.

Main Methods:

• A two-branch deep neural network combined with a multi-scale representation attention mechanism is proposed.
• WSIs are divided into bag-, patch-, and cell-level images, framing WSI classification as a MIL problem.
• The framework extracts features directly from all patches within each WSI, mining bag labels, significant patches, and cell-level information.

Main Results:

• The proposed framework demonstrates superior performance compared to state-of-the-art methods.
• Achieved higher classification accuracy in WSI analysis.
• Significantly improved model interpretability.

Conclusions:

• The novel deep MIL framework effectively addresses the challenges of gigapixel WSI analysis.
• The method offers a powerful tool for accurate and interpretable tumor diagnosis.
• The framework's ability to mine multi-level information enhances its diagnostic capabilities.