Focused active learning for histopathological image classification.

Area of Science:

• Digital Pathology
• Machine Learning
• Medical Imaging

Background:

• Efficiently acquiring labeled data is crucial for machine learning in digital pathology.
• Existing active learning (AL) methods struggle with artifacts, ambiguities, and class imbalance in medical data.
• Lack of precise uncertainty estimation leads to acquiring low-value images.

Purpose of the Study:

• To develop an active learning (AL) method that effectively acquires informative images for digital pathology.
• To address challenges posed by artifacts, ambiguities, and class imbalance in medical datasets.
• To improve the efficiency and performance of machine learning models in medical image analysis.

Main Methods:

• Proposed Focused Active Learning (FocAL), integrating a Bayesian Neural Network with Out-of-Distribution detection.
• FocAL estimates weighted epistemic uncertainty (for class imbalance), aleatoric uncertainty (for ambiguities), and an Out-of-Distribution score (for artifacts).
• Validated on MNIST and the real-world Panda dataset for prostate cancer classification.

Main Results:

• FocAL effectively focuses on informative images, avoiding ambiguities and artifacts that hinder other AL methods.
• Achieved a Cohen's kappa of 0.764 on the Panda dataset using only 0.69% of labeled data.
• Outperformed existing active learning approaches in both experimental settings.

Conclusions:

• FocAL enhances active learning by accurately estimating multiple uncertainty types for image acquisition.
• The method demonstrates superior performance and data efficiency in challenging medical imaging scenarios.
• FocAL offers a robust solution for the efficient labeling of digital pathology data.