Effective and efficient active learning for deep learning-based tissue image analysis.

Area of Science:

• Digital pathology
• Machine learning
• Computational biology

Background:

• Deep learning models require high-quality training datasets in digital pathology.
• Manual data annotation by expert pathologists is time-consuming and labor-intensive.
• Active learning (AL) aims to minimize annotation effort by selecting informative samples for labeling.

Purpose of the Study:

• To develop an optimized framework for active learning (AL) in digital pathology.
• To reduce annotation and execution time while improving model performance.
• To provide an integrated interface for domain specialists.

Main Methods:

• Developed a framework with a user interface and run-time optimizations for AL.
• Implemented diversity-aware data acquisition (DADA) to enhance data diversity and model performance.
• Utilized Network Auto-Reduction (NAR) for model simplification to decrease computational demands during AL.

Main Results:

• DADA demonstrated superior performance over state-of-the-art AL strategies for various convolutional neural networks (CNNs).
• Network Auto-Reduction (NAR) improved AL execution time by up to 4.3×.
• Models trained with data selected by NAR-reduced versions achieved comparable or better classification quality than those using full CNNs for selection.

Conclusions:

• The developed framework effectively reduces annotation and execution time in AL for digital pathology.
• The combination of DADA and NAR enhances model performance and computational efficiency.
• The integrated interface facilitates easier adoption and fine-tuning by domain specialists.