Attention induction based on pathologist annotations for improving whole slide pathology image classifier.

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in medicine

Background:

• Whole slide image (WSI) classification requires identifying diagnostically relevant regions.
• Attention mechanisms in multiple instance learning and hierarchical representation learning aim to focus on these regions.
• Training attention mechanisms with limited WSI data often results in suboptimal focus on informative areas.

Purpose of the Study:

• To propose and evaluate a novel attention induction method to enhance attention mechanisms in WSI classifiers.
• To improve the focus of attention mechanisms on diagnostically relevant regions for lesion classification.
• To boost WSI classification performance without requiring an increased number of training WSIs.

Main Methods:

• Developed an attention induction method tailored for hierarchical WSI representations.
• Utilized pathologist's coarse annotations to guide the attention mechanism.
• Integrated the attention induction method into existing WSI classification frameworks.

Main Results:

• The proposed attention induction method significantly improved the performance of the attention mechanism.
• Enhanced WSI classification accuracy was observed after applying the attention induction technique.
• The method effectively guided attention to focus on diagnostically relevant regions.

Conclusions:

• Attention induction is a viable strategy to enhance attention mechanisms in WSI classification.
• This method offers a solution for improving classifier performance with limited training data.
• The approach holds promise for improving automated analysis in digital pathology.