Multi-Faceted Consistency learning with active cross-labeling for barely-supervised 3D medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for 3D medical image segmentation requires extensive voxel-wise annotations, which are costly and time-consuming.
• Cross-annotation, using only a few orthogonal slices, offers a cost-effective alternative but poses challenges for barely-supervised learning (BSL).

Purpose of the Study:

• To develop an effective framework for active barely-supervised learning (BSL) in 3D medical image segmentation using minimal annotations.
• To enhance segmentation accuracy and generalizability despite sparse labeling.

Main Methods:

• Proposed a Multi-Faceted ConSistency learning (MF-ConS) framework combined with a Diversity and Uncertainty Sampling-based Active Learning (DUS-AL) strategy.
• Implemented a teacher-student architecture with three consistency regularization modules: neighbor-informed object prediction, prototype-driven consistency, and stability constraint.
• Utilized cross-annotation with only three orthogonal slices per scan and active learning to guide human annotation.

Main Results:

• The MF-ConS (DUS-AL) framework demonstrated superior performance compared to state-of-the-art methods on three benchmark datasets.
• Achieved consistent improvements in segmentation accuracy under extremely limited annotation conditions.
• The active learning strategy effectively directed annotations towards the most informative volumes.

Conclusions:

• The proposed MF-ConS framework with DUS-AL is highly effective for 3D medical image segmentation with minimal annotations.
• This approach addresses the challenges of BSL, improving object perception, feature compactness, and generalizability.
• Offers a practical solution for reducing annotation costs in medical imaging AI.