Active Learning Performance in Labeling Radiology Images Is 90% Effective.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Training artificial intelligence (AI) systems for radiology requires manual image labeling, a process that is time-consuming, costly, and prone to errors.
• Radiology image labeling involves human experts manually outlining regions and assigning labels, which is a bottleneck in developing large-scale AI models.

Purpose of the Study:

• To introduce and evaluate an active learning (AL) methodology to optimize the image labeling process for AI training in radiology.
• To demonstrate how AL can reduce the human effort required for labeling large radiology image datasets.

Main Methods:

• Implementing an active learning (AL) workflow that prioritizes the most informative unlabeled images for human review.
• Developing an advanced AL methodology by introducing five distinct elements to the standard AL process.
• Validating the effectiveness of the proposed AL approach using three real-life radiology datasets.

Main Results:

• Active learning significantly reduces the need for manual labeling, with most information learned after approximately 10% of the dataset is labeled.
• The majority of remaining images can be automatically labeled and subsequently verified by a radiologist, drastically decreasing overall human effort.
• The advanced AL methodology proved effective across diverse, real-world radiology datasets.

Conclusions:

• Active learning is a highly effective strategy for streamlining the labeling of radiology images for AI development.
• The proposed advanced AL methodology offers a substantial improvement over standard AL, leading to more efficient and cost-effective AI model training.
• This approach has the potential to accelerate the deployment of AI in medical imaging by reducing the data annotation bottleneck.