The inherent ambiguity in distinguishing boundaries between lesions and adjacent tissues makes accurate segmentation in ultrasound images challenging. Although deep learning has improved segmentation accuracy, boundary segmentation quality and its relationship with anatomical structures remain underexplored. To address this, we propose DBF-Net, a dual-branch deep neural network that captures supervised relationships between anatomical structures and boundaries. Additionally, we introduce a feature fusion module to enhance the integration of body and boundary information. We evaluate DBF-Net on three public ultrasound image datasets and demonstrate its superiority over existing methods, achieving Dice Similarity Coefficients of 81.05±10.44% for breast cancer, 76.41±5.52% for brachial plexus nerves, and 87.75±4.18% for infantile hemangiomas on the BUSI, UNS, and UHES datasets, respectively. Our approach outperforms current methods, showcasing its effectiveness in ultrasound image segmentation. Code available at: https://github.com/apple1986/DBF-Net.

A sharp prediction for ultrasound image segmentation is critical to assist doctors in the accurate diagnosis and treatment. Although the existing methods could achieve impressive performance on biomedical image segmentation, they still have difficulty in segmenting pixels near the boundaries of lesions faithfully. To this end, we propose a body and boundary supervision network aiming to improve ultrasound image segmentation performance, especially on the boundary of the segmented objects. In contrast to existing approaches that take lesions as a whole, a novel dual-branch supervision block was developed to explore explicit modeling of the body and boundary of the object at feature level. Furthermore, we presented an adaptive feature fusion strategy to combine feature maps from dual-branch to get more representative features for final segmentation. Experimental results on three public ultrasound datasets including BUSI, HC18 and Hemangioma demonstrate that the proposed network can achieve leading segmentation performance for breast cancer, fetal head circumference and hemangioma on ultrasound images compared with the existing state-of-the-art models.

We present a highly efficient method for interactive volumetric meshless shape deformation. Our method operates within a low dimensional sub-space of shape-aware C ∞ harmonic maps, and is the first method that is guaranteed to produce a smooth locally injective deformation in 3D. Unlike mesh-based methods in which local injectivity is enforced on tetrahedral elements, our method enforces injectivity on a sparse set of domain samples. The main difficulty is then to certify the map as locally injective throughout the entire domain. This is done by utilizing the Lipschitz continuity property of the harmonic basis functions. We show a surprising relation between the Lipschitz constant of the smallest singular value of the map Jacobian and the norm of the Hessian. We further carefully derive a Lipschitz constant for the Hessian, and develop a sufficient condition for the injectivity certification. This is done by utilizing the special structure of the harmonic basis functions combined with a novel regularization term that pushes the Lipschitz constants further down. As a result, the injectivity analysis can be performed on a relatively sparse set of samples. Combined with a parallel GPU-based implementation, our method can produce superior deformations with unique quality guarantees at real-time rates which were possible only in 2D so far.