Scalable histopathological image analysis via active learning.

Area of Science:

• * Computational pathology and medical image analysis.
• * Machine learning and artificial intelligence in healthcare.

Background:

• * Medical image analysis requires large labeled datasets, posing a significant annotation burden on pathologists.
• * Active learning (AL) methods reduce labeling costs but often overlook structured information within medical data, such as patient-specific details.
• * Existing AL methods assume data independence, which may not hold true for real-world medical imaging datasets.

Purpose of the Study:

• * To develop a novel batch-mode active learning method for medical image analysis.
• * To leverage the inherent structured information in medical images to enhance data selection diversity.
• * To maximize information gain for improved model training with reduced annotation effort.

Main Methods:

• * Formulation of the active learning problem as an adaptive submodular function maximization.
• * Inclusion of a partition matroid constraint to enforce diversity in selected data batches.
• * Development of an efficient greedy algorithm with a theoretically proven performance bound.

Main Results:

• * The proposed method effectively utilizes structured information in medical image annotations.
• * Demonstrated significant reduction in annotation burden while maintaining predictive accuracy.
• * Successful application on a large dataset of thousands of breast microscopic tissue histopathological images.

Conclusions:

• * The novel batch-mode active learning approach enhances information gain by leveraging structured data.
• * This method offers a scalable and efficient solution for training medical image analysis systems.
• * The findings have implications for reducing pathologist workload and improving diagnostic AI accuracy.