.png){kind=icon}
.png){kind=icon}
Return
DBM-SAM: Dual-branch multiscale adaptation of SAM for medical ultrasound segmentation
.png){kind=icon}
PRE
AI
W
L
C
S
代
DOI：10.1016/j.displa.2026.103446.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Medical ultrasound imaging is widely used due to its non-invasive, real-time, portable, and safe characteristics, supporting tumor screening and lesion monitoring in organs such as the breast, thyroid, and kidney. Precise ultrasound segmentation is essential for quantifying lesion extent, assessing tumor progression, and improving diagnostic consistency. However, ultrasound images are prone to speckle noise, low contrast, shadow artifacts, and blurred boundaries, posing significant challenges for automated segmentation. Motivated by these challenges, we propose DBM-SAM, a dual-branch framework integrating a Convolutional Neural Network (CNN) branch for local features and a Vision Transformer (ViT) branch for long-range dependencies. Specifically, we introduce a compact CNN encoder equipped with a Multiscale Perception Module (MPM) to extract local features across different resolutions and adapt the ViT encoder with a lightweight Multi-cognitive Visual Adapter (Mona) to improve global contextual feature extraction. Furthermore, a Medical Mask Decoder is designed to replace the original multi-mask strategy in SAM with a single deterministic output, aiming to reduce ambiguity and better meet clinical requirements. Our method achieved average Recall, Precision, IoU, Dice, and 95% Hausdorff Distance (HD95) scores of 88.49%, 88.59%, 75.70%, 85.75%, and 36.25 mm, respectively, on two internal datasets (TN3K and BUSI). Evaluation on three external datasets (DDTI, DatasetB, and KUS) obtained average Recall, Precision, IoU, Dice, and HD95 scores of 86.24%, 87.49%, 73.48%, 83.61%, and 33.01 mm, respectively. These results demonstrate that the proposed approach exhibits robust performance and strong generalization capabilities in ultrasound image segmentation.
Keywords:
Ultrasound image
Segmentation
SAM
Dual-branch framework
Multiscale perception
Journal
IF:
3.4
Papers:
2.1K
Citations:
3.2K
Organization
Cited Papers