Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential

Area of Science:

• Digital pathology
• Computer-assisted diagnosis
• Machine learning

Background:

• Digital whole slide images offer vast data for automated analysis.
• Deep neural networks show promise in digital pathology tasks.
• Training deep learning models typically requires extensive manual annotations.

Purpose of the Study:

• To provide an overview of multiple instance learning (MIL) approaches in deep learning for digital pathology.
• To discuss recent advancements and challenges in MIL for whole slide image analysis.
• To highlight the potential of MIL in overcoming annotation limitations in digital pathology.

Main Methods:

• Review of widely used deep multiple instance learning concepts.
• Analysis of recent advancements in MIL algorithms.
• Discussion of techniques applicable to whole slide image analysis without full annotation.

Main Results:

• Multiple instance learning (MIL) is a powerful tool for training deep neural networks without fully annotated data.
• MIL is particularly effective in digital pathology where whole slide image labels are available but patch/pixel labels are not.
• Significant growth in MIL publications for digital pathology observed in recent years.

Conclusions:

• Deep multiple instance learning (MIL) presents a viable solution for training models on large-scale digital pathology data.
• The effectiveness of MIL is driven by data availability and advancements in GPU computing.
• Future research should address remaining challenges and explore further potential of MIL in digital pathology.