Proportion constrained weakly supervised histopathology image classification.

Area of Science:

• Computational pathology
• Machine learning
• Computer vision

Background:

• Multiple instance learning (MIL) is a weakly supervised learning paradigm popular in histological image analysis.
• MIL alleviates detailed labeling of large Whole Slide Images (WSIs) but often requires large datasets and struggles with multi-label classification and constraints.
• Existing MIL methods often focus on binary classification, which is insufficient for complex real-world problems with multi-label settings.

Purpose of the Study:

• To propose a novel multi-label MIL formulation incorporating inequality constraints to integrate prior knowledge of instance proportions.
• To provide a theoretical foundation based on optimization with log-barrier extensions for bag-level class proportions.
• To improve the performance of MIL in histological image analysis, particularly for multi-label classification tasks with constraints.

Main Methods:

• Developed a novel multi-label MIL formulation utilizing inequality constraints.
• Applied optimization with log-barrier extensions to model bag-level class proportions.
• Incorporated prior knowledge about instance proportions into the MIL training process.

Main Results:

• Achieved instance-level results comparable to supervised methods on datasets of similar size using prior proportion information.
• Demonstrated approximately 13% improvement in instance-level accuracy compared to prior MIL settings.
• Showcased approximately 3% improvement in the multi-label mean area under the ROC curve at the bag-level.

Conclusions:

• The proposed MIL method effectively incorporates prior proportion information, enhancing performance in histological image analysis.
• The method shows significant improvements in both instance-level accuracy and bag-level multi-label classification.
• This approach offers a more robust solution for complex, real-world pathological image analysis tasks.