A Novel Attribute-Based Symmetric Multiple Instance Learning for Histopathological Image Analysis

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Histopathological image analysis faces challenges due to diverse features and large non-informative regions in whole slide images.
• Existing multiple-instance learning (MIL) methods often rely on asymmetric assumptions about instance relevance within bags, which may not suit all scenarios.

Purpose of the Study:

• To propose a novel symmetric MIL framework for accurate image-level classification and region annotation in histopathology.
• To address the limitations of traditional MIL by accommodating representative negative instances in negative bags.

Main Methods:

• Developed a symmetric MIL framework where each instance is assigned an attribute: negative, positive, or irrelevant.
• Introduced a probabilistic graphical model with controlled relevance and efficient inference for parameter learning.
• Created an instance-level attribute-learning classifier.

Main Results:

• The proposed symmetric MIL framework demonstrated effectiveness in histopathological image analysis.
• Achieved promising results on available histopathology datasets for both classification and annotation tasks.

Conclusions:

• The novel symmetric MIL approach offers a more flexible and accurate solution for histopathological image analysis.
• This method enhances the ability to classify whole slide images and identify diagnostically relevant regions.