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

Ultrasound image segmentation is often limited by the scarcity of annotated datasets, especially in resource-constrained clinical settings. To address this issue, we employ BT-UNet, a self-supervised learning framework that combines Barlow Twins (BT) with the UNet architecture, and aim to enhance segmentation performance in low-data conditions. Unlike previous work that trains BT-UNet exclusively on clinical datasets, our approach explores the benefits of pre-training BT-UNet on musculoskeletal phantom ultrasound images, before fine-tuning it on a small set of annotated clinical images. Our results demonstrate that this strategy significantly improves segmentation performance under limited annotated data. Specifically, with only 5% of the labeled clinical dataset, BT-UNet achieves a Dice score of 0.9311, slightly outperforming the standard UNet's 0.9250. However, at an extreme data scarcity level of 1%, BT-UNet maintains a Dice score of 0.7114, whereas UNet drops to 0.2253. These results highlight the potential of self-supervised pre-training on phantom datasets to address data scarcity challenges in medical imaging. By utilizing unlabeled phantom data for representation learning, BT-UNet enhances segmentation accuracy with minimal clinical annotations, offering a promising solution for real-world medical applications where annotated data is limited.Clinical relevance: This study shows that pre-training a self-supervised learning model on musculoskeletal phantom ultrasound images and fine-tuning it with limited clinical data can significantly improve segmentation accuracy, offering a promising solution to reduce reliance on large annotated datasets.