Ensemble Distillation of Divergent Opinions for Robust Pathological Image Classification.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Accurate deep neural networks (DNNs) rely on consistent, labeled data.
• Inter-observer variability in labeling, especially in pathological image diagnosis, poses a significant challenge.
• Existing studies on building DNNs with ambiguous ground truth are limited.

Purpose of the Study:

• To address the challenge of inter-observer variability in pathological image diagnosis.
• To develop a method for constructing robust DNN models despite inconsistent expert labels.
• To improve the generalization capability and classification accuracy of DNNs in the presence of labeling ambiguities.

Main Methods:

• Proposed a novel method for constructing DNN models that leverages knowledge from multiple DNNs.
• Exploited relationships among data learned by diverse DNN models to enhance robustness.
• Conducted comparative experiments using multiple pathology datasets with independent labeling by different pathologists.

Main Results:

• The proposed method demonstrated good generalization capability across different datasets.
• Achieved superior classification accuracy compared to baseline models.
• Effectively mitigated the impact of inter-observer variability on DNN performance.

Conclusions:

• The developed method offers a robust approach to building DNNs for pathological image diagnosis with inter-observer variability.
• This technique enhances the reliability and accuracy of AI-driven diagnostic tools.
• Addresses a critical gap in applying DNNs to subjective medical image interpretation.