Label-Efficient Breast Cancer Histopathological Image Classification.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Computational Pathology

Background:

• Accurate breast cancer classification from histopathological images is crucial for diagnosis.
• Deep learning models require large labeled datasets, which are costly and time-consuming to create.
• Current methods face challenges in efficient data annotation for histopathological analysis.

Purpose of the Study:

• To develop a deep active learning framework for breast cancer histopathological image classification.
• To maximize learning accuracy with minimal manual labeling efforts.
• To reduce the cost and time associated with expert annotation.

Main Methods:

• Implemented a deep active learning framework for image classification.
• Utilized manual annotation of the most informative unlabeled samples.
• Incorporated iterative model updates with an expanding training set.
• Evaluated two sample selection strategies: entropy-based and confidence-boosting.

Main Results:

• The proposed framework significantly reduces annotation costs by up to 66.67% compared to random selection.
• Achieved higher classification accuracy with reduced annotation effort.
• Demonstrated superior performance over standard query strategies in terms of cost and accuracy.

Conclusions:

• Deep active learning is an effective approach for breast cancer histopathological image classification.
• The proposed framework offers a cost-effective and efficient solution for training deep learning models.
• Intelligent sample selection is key to maximizing performance with limited labeled data.