Effective active learning in digital pathology: A case study in tumor infiltrating lymphocytes

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in medicine

Background:

• Deep learning excels in pathology image analysis but demands extensive expert-annotated data.
• Minimizing annotation needs is crucial for broader adoption of deep learning in pathology.

Purpose of the Study:

• To reduce the data annotation requirements for deep learning in pathology image analysis.
• To develop and evaluate an efficient active learning (AL) strategy for pathology tasks.

Main Methods:

• Implemented an iterative active learning (AL) approach for Tumor Infiltrating Lymphocytes (TIL) classification.
• Proposed a novel AL acquisition method utilizing data grouping by imaging features and prediction uncertainty.
• Evaluated state-of-the-art AL methods alongside the proposed strategy.

Main Results:

• The proposed method significantly reduced the number of image patches needed to achieve a target Area Under the Curve (AUC).
• An optimization technique (subpooling) resulted in approximately a 2.12× improvement in AL execution time.
• Experimental validation on cancer tissue images confirmed the efficacy of the developed strategy.

Conclusions:

• The strategy enables deep learning analyses for TIL classification with substantially reduced annotation effort.
• This approach holds potential for developing other digital pathology analyses requiring fewer training samples.
• The findings contribute to making advanced AI tools more accessible in pathology research and practice.