Active learning for medical image segmentation with stochastic batches.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• High-quality labelled data is crucial for machine learning algorithm performance, but manual annotation for medical image segmentation is resource-intensive.
• Active learning (AL) aims to reduce annotation effort by intelligently selecting informative samples for expert labelling.
• Existing AL methods often struggle with medical image segmentation due to limitations in uncertainty quantification and computational costs.

Purpose of the Study:

• To enhance uncertainty-based active learning strategies for medical image segmentation.
• To address the sub-optimal batch-querying strategies in current uncertainty-based AL methods.
• To develop a computationally efficient and effective AL approach for medical image segmentation.

Main Methods:

• Proposed a novel approach using stochastic batches (SB) to compute uncertainty at the batch level, rather than the sample level.
• Integrated stochastic batch querying as an add-on to existing uncertainty-based AL metrics.
• Evaluated the method extensively on two distinct medical image segmentation datasets.

Main Results:

• The proposed stochastic batch querying strategy consistently outperformed conventional uncertainty-based sampling methods.
• The method demonstrated improved performance in selecting informative samples for training medical image segmentation models.
• Achieved a strong baseline performance for active learning in the domain of medical image segmentation.

Conclusions:

• Computing uncertainty at the batch level using stochastic batches is an effective strategy for improving active learning in medical image segmentation.
• The proposed method offers a simple yet powerful enhancement to existing uncertainty-based AL techniques.
• This work provides a robust and efficient baseline for future research in medical image segmentation annotation.