SingleStrip: learning skull-stripping from a single labeled example.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning segmentation for volumetric images like brain MRI demands extensive labeled data, a significant bottleneck.
• Existing domain randomization methods have limited anatomical variability with scarce labeled data.
• Semi-supervised self-training leverages unlabeled data to overcome label scarcity.

Purpose of the Study:

• To develop a novel semi-supervised deep learning strategy for 3D skull-stripping using minimal labeled brain MRI data.
• To combine domain randomization with self-training for enhanced segmentation performance.
• To evaluate an autoencoder-based quality control method for pseudo-label selection.

Main Methods:

• Domain randomization was used to synthesize training images from limited label maps.
• A convolutional autoencoder (AE) was trained on a single labeled example for quality assessment of predicted masks.
• Pseudo-labels ranked by AE reconstruction error were used to fine-tune the skull-stripping network.

Main Results:

• The combined approach enabled effective 3D skull-stripping with as few as one labeled example.
• Performance on out-of-distribution data approached that of models trained with more labeled data.
• AE-based quality control showed stronger correlation with segmentation accuracy than consistency-based ranking.

Conclusions:

• Combining domain randomization and AE-based quality control facilitates effective semi-supervised segmentation from extremely limited labeled data.
• This strategy can significantly reduce the labeling burden in medical imaging studies.
• The method shows promise for new anatomical structures and emerging imaging techniques.