Masked pretraining of U-Net for ultrasound image segmentation

Le résumé fourni par la source

Computer-aided segmentation of ultrasound images can assist less experienced sonographers in making a diagnosis or performing early screening for diseases in vast rural areas. U-Net is the mainstream method for ultrasound image segmentation, offering advantages of efficient feature extraction and context integration. However, existing U-Net-based segmentation methods can only yield modest performance due to the low quality of ultrasound imaging and the usually small amount of available labeled data. This research proposes a self-supervised pretraining method to promote the performance of U-Net-based models for ultrasound image segmentation, when limited labeled images but numerous relevant unlabeled ones are available, which is a common situation in ultrasonography scenarios such as rare disease detection. By randomly masking out part of pixels in the input ultrasound image, the U-Net based model is pretrained to predict the unknown content, to adapt the model to the current fitting scenario before the formal training. A lightweight but effective variant of U-Net named MS-UNet is also proposed to better fit the scenario of ultrasound image segmentation. Experimental results show that the masked pretraining can boost the segmentation performance of U-Net models on small-size ultrasound image datasets, with Dice score improvements of 6-20 percentage points across four datasets under minimal labeled data conditions. Furthermore, our proposed MS-UNet achieves a relatively high segmentation accuracy while reducing computational complexity by 80.3% in FLOPs and 53.2% in parameters compared to the standard U-Net.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.