Collaborative instance-level and bag-level multiple instance learning with label disambiguation for whole slide image...

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in healthcare

Background:

• Artificial intelligence (AI) is transforming histopathological image analysis for disease diagnosis and treatment.
• Whole slide images (WSIs) present analysis challenges due to their gigapixel size, necessitating multiple instance learning (MIL) methods.
• Current MIL methods struggle with noisy pseudo-labels (instance-level) and suboptimal feature aggregation (bag-level).

Purpose of the Study:

• To develop a novel MIL method, CIB-MIL, for WSI analysis.
• To address limitations of existing MIL approaches by integrating collaborative instance-level and bag-level supervision.
• To improve the accuracy and robustness of AI-driven histopathological image analysis.

Main Methods:

• Introduced CIB-MIL, a novel MIL method for WSI analysis.
• Implemented a label disambiguation module with noisy-label learning for refining instance pseudo-labels.
• Developed a collaborative supervision framework enabling interaction between attention and pseudo-label mechanisms for cooperative optimization.

Main Results:

• CIB-MIL demonstrated state-of-the-art performance across five diverse datasets (three public, two in-house).
• The method effectively mitigated noisy labels and improved feature aggregation through collaborative supervision.
• Achieved superior results in histopathological image analysis tasks.

Conclusions:

• CIB-MIL offers a significant advancement in AI-driven WSI analysis.
• The collaborative supervision approach enhances diagnostic accuracy and robustness.
• This method holds promise for improving disease diagnosis, prognosis, and treatment planning in healthcare.