HAL-IA: A Hybrid Active Learning framework using Interactive Annotation for medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for medical image segmentation requires extensive pixel-wise annotations, which are time-consuming and expensive to acquire.
• Reducing annotation costs is critical for the widespread clinical application of medical image segmentation.

Purpose of the Study:

• To develop an efficient framework for medical image segmentation that minimizes annotation costs.
• To address the challenges of cold start, sample selection, and manual annotation burden in active learning for segmentation.

Main Methods:

• Proposed a Hybrid Active Learning framework using Interactive Annotation (HAL-IA).
• Introduced a novel hybrid sample selection strategy combining pixel entropy, regional consistency, and image diversity.
• Implemented a warm-start initialization strategy to overcome the cold-start problem.
• Developed an interactive annotation module with suggested superpixels for simplified labeling.

Main Results:

• The HAL-IA framework achieved high accuracy in pixel-wise annotations and segmentation models.
• Demonstrated significant reduction in labeled data requirements and manual interactions compared to existing methods.
• Outperformed state-of-the-art methods across four medical image datasets.

Conclusions:

• HAL-IA effectively reduces annotation costs for medical image segmentation while maintaining high accuracy.
• The framework offers a practical solution for physicians to obtain accurate segmentation results for clinical analysis and diagnosis.
• This approach facilitates efficient medical image analysis and supports improved patient care.