A Multi-scale Linear-time Encoder for Whole-Slide Image Analysis.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Biomedical image analysis

Background:

• Whole-slide image (WSI) analysis is complex due to gigapixel sizes and multiple magnifications.
• Existing multiple instance learning (MIL) methods often use single-scale analysis.
• Transformer models face computational challenges with quadratic attention costs.

Purpose of the Study:

• Introduce MARBLE, a purely Mamba-based MIL framework for efficient WSI analysis.
• Enable parallel processing of multiple magnification levels for integrated reasoning.
• Provide a scalable and modular alternative to attention-based architectures.

Main Methods:

• Developed the Multi-scale Adaptive Recurrent Biomedical Linear-time Encoder (MARBLE).
• Employed a Mamba-based state-space model for linear-time sequence modeling.
• Implemented parallel processing of multiple WSI scales with coarse-to-fine reasoning.

Main Results:

• MARBLE demonstrated significant performance improvements across five public datasets.
• Achieved up to 6.9% increase in Area Under the Curve (AUC).
• Showcased up to 20.3% improvement in accuracy and 2.3% in C-index.

Conclusions:

• MARBLE establishes a new standard for efficient and generalizable multi-scale WSI analysis.
• The Mamba-based approach offers a scalable solution overcoming limitations of existing methods.
• MARBLE effectively captures cross-scale dependencies with minimal parameter overhead.