Improving Ultrasound Image Segmentation in Data-Scarce Scenarios Using Self-Supervised Learning With Phantom Data...

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Ultrasound image segmentation is crucial but hindered by limited annotated datasets, especially in clinical settings.
• Existing methods struggle with data scarcity, impacting diagnostic accuracy.

Purpose of the Study:

• To enhance ultrasound image segmentation performance in low-data conditions using a self-supervised learning framework.
• To investigate the efficacy of pre-training on phantom ultrasound data before fine-tuning on clinical data.

Main Methods:

• Employed BT-UNet, a self-supervised framework combining Barlow Twins (BT) with UNet architecture.
• Pre-trained BT-UNet on unlabeled musculoskeletal phantom ultrasound images.
• Fine-tuned the model on a small set of annotated clinical ultrasound images.

Main Results:

• BT-UNet achieved a Dice score of 0.9311 with 5% labeled clinical data, surpassing standard UNet (0.9250).
• At 1% data scarcity, BT-UNet maintained a Dice score of 0.7114, while UNet dropped to 0.2253.
• Demonstrated significant performance improvement under extreme data scarcity.

Conclusions:

• Self-supervised pre-training on phantom datasets effectively addresses data scarcity in medical imaging segmentation.
• BT-UNet offers a promising solution for improving segmentation accuracy with minimal clinical annotations.
• Reduces reliance on large, costly annotated datasets for clinical applications.