Diminishing Uncertainty Within the Training Pool: Active Learning for Medical Image Segmentation

Area of Science:

• Machine Learning
• Medical Image Analysis
• Computer Vision

Background:

• Active learning (AL) is a machine learning paradigm where algorithms select data for annotation, unlike passive learning.
• Existing AL frameworks primarily focus on classification tasks, with limited exploration in medical image segmentation.
• Efficiently annotating large medical imaging datasets is a significant bottleneck in developing accurate AI models.

Purpose of the Study:

• To investigate the efficacy of active learning for medical image segmentation tasks.
• To propose and evaluate novel active learning strategies tailored for segmentation.
• To demonstrate significant data reduction in training segmentation models for medical imaging.

Main Methods:

• A query-by-committee approach for active learning with a joint optimizer.
• Three novel AL strategies: increasing uncertain data frequency, using mutual information for diversity, and adapting Dice log-likelihood for SVGD.
• Validation on two medical imaging datasets: MRI scans of the hippocampus and CT scans of pancreas/tumors.

Main Results:

• The proposed active learning framework achieved full accuracy in segmentation tasks.
• Significant data reduction was observed, utilizing only 22.69% of MRI data and 48.85% of CT data.
• The novel strategies contributed to efficient learning and improved model performance with less data.

Conclusions:

• Active learning is highly effective for accelerating the training of medical image segmentation models.
• The proposed framework and strategies substantially reduce the annotation effort required for high-performance models.
• This approach offers a promising solution for overcoming data scarcity challenges in medical AI.