Self-supervision enhances instance-based multiple instance learning methods in digital pathology: a benchmark study.

Area of Science:

• Computational pathology
• Machine learning for medical imaging

Background:

• Multiple instance learning (MIL) is a key technique for whole slide image (WSI) classification, treating slides as bags of patches.
• MIL methods are broadly categorized into instance-based and embedding-based approaches.
• Historically, embedding-based MIL dominated due to robustness, despite instance-based methods offering better interpretability.

Purpose of the Study:

• To evaluate the performance of instance-based versus embedding-based MIL strategies in WSI classification.
• To investigate the impact of recent advancements in self-supervised learning (SSL) on MIL performance.
• To introduce novel instance-based MIL methods tailored for the pathology domain.

Main Methods:

• Conducted 710 experiments across 4 datasets.
• Compared 10 MIL strategies, 6 SSL methods with 4 backbones, and 4 foundation models.
• Introduced 4 new instance-based MIL methods for pathology applications.

Main Results:

• Instance-based MIL methods with effective SSL feature extractors match or surpass complex embedding-based methods.
• New state-of-the-art results were achieved on the BRACS and Camelyon16 datasets.
• Simple instance-based MIL models demonstrated strong performance with fewer parameters.

Conclusions:

• Advanced SSL techniques enhance the performance of simple instance-based MIL for WSI classification.
• Instance-based MIL methods offer superior interpretability for clinical applications.
• Future research should prioritize developing tailored SSL methods for WSI over complex embedding-based MIL approaches.