UniSAL: Unified Semi-supervised Active Learning for histopathological image classification.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Machine learning for medical imaging

Background:

• Histopathological image classification is vital for cancer diagnosis.
• Deep learning models require large annotated datasets, which are expensive and time-consuming to create.
• Limited labeled data hinders the training of effective deep neural networks for pathology.

Purpose of the Study:

• To develop an efficient annotation framework, Unified Semi-supervised Active Learning (UniSAL), to reduce human effort.
• To improve the efficiency of sample selection for annotation in histopathological image classification.
• To achieve high performance in cancer diagnosis with significantly reduced annotation costs.

Main Methods:

• Proposed a Unified Semi-supervised Active Learning (UniSAL) framework for selecting informative and representative samples.
• Introduced dual-view high-confidence pseudo training utilizing both labeled and unlabeled images.
• Implemented pseudo label-guided class-wise contrastive learning for enhanced feature representation.
• Designed a Disagreement-aware Uncertainty Selector (DUS) and a Compact Selector (CS) for sample selection.

Main Results:

• UniSAL significantly outperformed state-of-the-art active learning methods on three public datasets (CRC5000, Chaoyang, CRC100K).
• The framework reduced annotation costs to approximately 10%.
• Achieved performance comparable to full annotation with the reduced dataset.

Conclusions:

• UniSAL offers an effective solution for reducing annotation costs in histopathological image classification.
• The proposed framework enables efficient training of deep learning models for cancer diagnosis.
• Dual-view pseudo training and novel sample selection strategies are key to UniSAL's success.