An active learning based classification strategy for the minority class problem: application to histopathology...

Area of Science:

• Digital pathology
• Machine learning in healthcare
• Computational biology

Background:

• Supervised classifiers in digital pathology aid disease detection but require expert-labeled training data.
• Digital pathology datasets often exhibit a

Purpose of the Study:

• To develop an effective training strategy for supervised classifiers in digital pathology.
• To address the challenges of limited labeled data and the minority class problem in biomedical image analysis.

Main Methods:

• A novel training strategy combining active learning (AL) with class-balancing (CBAL) was developed.
• A mathematical model was created to predict the annotation cost for achieving balanced training classes.
• The CBAL strategy was applied to classify cancer in digitized prostate histopathology images.

Main Results:

• The CBAL-trained classifier demonstrated superior performance, achieving 2% greater accuracy and 3% higher AUC compared to unbalanced AL and random learning methods.
• The developed cost model accurately predicted the number of annotations required for balanced classes.
• Over-sampling the minority class provided marginal accuracy gains at a higher annotation cost.

Conclusions:

• The combined AL and class-balancing strategy offers a generalizable solution for supervised classification problems with expensive datasets and minority class issues.
• This intelligent training approach, integrating AL, class ratio optimization, and cost modeling, enhances the efficiency and effectiveness of the training process for researchers.