Learning From Incorrectness: Active Learning With Negative Pre-Training and Curriculum Querying for Histological...

Area of Science:

• Digital pathology
• Computational biology
• Machine learning in medicine

Background:

• Histological tissue classification is crucial for slide analysis but demands high annotation costs for deep learning.
• Active learning (AL) offers a solution to reduce annotation budgets in this domain.
• Existing AL methods struggle with performance imbalance across categories, impacting diagnostic accuracy.

Purpose of the Study:

• To develop an active learning framework (ICAL) that addresses performance imbalance and reduces annotation costs for histological tissue classification.
• To improve the accuracy and balance of classification for all tissue categories, including those with insufficient performance.
• To enable high-performance histological analysis with significantly reduced labeled data.

Main Methods:

• Proposed ICAL framework with Incorrectness Negative Pre-training (INP) and Category-wise Curriculum Querying (CCQ).
• INP uses incorrect predictions as complementary labels for negative pre-training to distinguish similar categories.
• CCQ adjusts query weights based on category learning status and uses uncertainty to mitigate bias.

Main Results:

• ICAL achieved performance close to fully supervised learning using less than 16% of labeled data.
• Demonstrated superior and more balanced performance across all categories compared to state-of-the-art AL algorithms.
• Maintained robustness even with extremely low annotation budgets on two histological datasets.

Conclusions:

• ICAL effectively reduces annotation costs for histological tissue classification while maintaining high performance.
• The proposed framework ensures balanced performance across all categories, crucial for accurate cancer diagnosis.
• ICAL represents a significant advancement in efficient and robust histological slide analysis using active learning.