Segmentation of Ultrasound Images Using Customized Finetuning Strategies Based on Foundation Model

Deep learning as a typical data-driven technique has shown great potential in computer-aided ultrasound image segmentation, but how to train deep models with limited data samples is still a challenge. In this study, we developed a method to transfer the pre-trained foundation model to ultrasound image segmentation and utilized a particular training strategy to train the prompt-based model not being limited by prompts. We customize strategies that combine appropriate finetune approaches to broaden the scope of ultrasound image segmentation, despite limited datasets and cost constraints. By introducing convolutional bypasses in the image encoder to extract features and finetuning the mask decoder at the same time, the segmentation performance has been further improved while the number of training parameters is reduced to less than 5% of the original network. The proposed method achieves the highest DSC of 89.12%, 89.78%, 84.49% and 95.18% on BUSI, DDTI, TN3K and CAMUS datasets with proper prompt inputs, containing three different organs: breast, thyroid, and heart, respectively. Compared with those SOTA segmentation networks, the proposed method also demonstrated comparable segmentation capabilities without prompts. It is expected to be a powerful tool for diverse lesion segmentation of ultrasound images.