Self-Supervised Learning to More Efficiently Generate Segmentation Masks for Wrist Ultrasound

Deep learning automation of medical image analysis is highly desirable for purposes including organ/tissue segmentation and disease detection. However, deep learning traditionally relies on supervised training methods, while medical images are far more expensive to label than natural images. Self-supervised learning (SSL) has been gaining attention as a technique that allows strong model performance with only a small amount of labeled data. This would be particularly useful in ultrasound (US) imaging, which can involve hundreds of images per video sweep, saving time and money for labeling. In this paper, we proposed a new SSL-based image segmentation technique that can be applied to bone segmentation in wrist US. This is the first use of the classification models SSL pretraining method SimMIM in wrist US. We modified the SimMIM SSL pretraining architecture, used a speckle noise masking policy to generate noise artifacts similar to those seen in US, changed the loss function, and analyzed how they influenced the downstream segmentation tasks. Using modified SimMIM, our approach surpassed the performance of state-of-the-art fully supervised models on wrist bony region segmentation by up to 3.2% higher Dice score and up to 4.5% higher Jaccard index, using an extremely small labeled dataset with only 187/935 images and generated labels visually consistent with human labeling on the test set of 3822 images. The SSL pretrained models were also robust on the test set annotated by different medical experts.