Targeting tumor heterogeneity: multiplex-detection-based multiple instance learning for whole slide image...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Machine learning for medical imaging

Background:

• Multiple instance learning (MIL) is crucial for classifying whole slide images (WSIs) in diagnostic pathology.
• Identifying critical instances within WSIs is challenging due to tumor heterogeneity, which often hinders algorithm performance.

Purpose of the Study:

• To propose a novel multiplex-detection-based multiple instance learning (MDMIL) method to effectively target and overcome tumor heterogeneity in WSI classification.
• To enhance the instance-mining capacity and robustness of deep learning models for computational pathology.

Main Methods:

• Developed a novel MDMIL framework utilizing multiplex detection strategies with internal and variational queries for improved instance identification.
• Introduced a memory-based contrastive loss function to ensure feature space consistency across diverse phenotypes.
• Implemented feature constraints among samples to bolster the network's ability to handle tumor heterogeneity.

Main Results:

• The proposed MDMIL approach demonstrated significantly improved instance-mining capacity.
• The novel network and loss function achieved high robustness against tumor heterogeneity.
• Experiments on CAMELYON16, TCGA-NSCLC, and TCGA-RCC datasets showed MDMIL outperformed existing state-of-the-art methods.

Conclusions:

• MDMIL effectively addresses the challenge of tumor heterogeneity in WSI classification.
• The method offers a robust and high-performing solution for diagnostic pathology applications.
• The developed MDMIL framework provides a valuable tool for computational pathology research.