Foundation-guided active learning for efficient ultrasound image segmentation

Medical image segmentation plays a vital role in computer-aided diagnosis and treatment planning. Despite the success of deep learning models, particularly U-Net-based architectures, their performance heavily depends on the availability of large, well-annotated datasets, which are often scarce due to the high cost and expertise required for annotation. Active Learning (AL) presents a promising approach to alleviate the annotation burden by selecting the most informative samples for labeling. However, traditional AL approaches suffer from cold start issues in the early rounds due to the insufficient amount of labeled data. In this work, we propose a foundation-guided deep active learning framework for ultrasound image segmentation. By leveraging pre-trained foundation models, such as DINOv2, we extract rich feature representations that improve both sample selection and model performance, especially during the initial AL rounds. Our approach integrates these models into a U-Net-based segmentation pipeline, demonstrating enhanced data efficiency and segmentation accuracy. By combining the generalization capability of foundation models with task-specific training, our framework provides a practical solution for scalable, label-efficient ultrasound image segmentation. Our code is available at https: //github.com/trnKhanh/medical-image-analysis