Comprehensive Summary
The study proposes a lightweight segmentation framework called MCBL-UNet which combines the long-range modeling capability of Mamba layers with convolutional neural networks (CNNs) for efficient placenta ultrasound image segmentation. It uses a compact 6-layer U-Net architecture with several key modules: a boundary enhancement module (BEM), a multi-dimensional global context module (MGCM), a parallel channel spatial attention module (PCSAM), an attention downsampling module (ADM) and a content-aware upsampling module (CUM). The model only has ~1.31 million parameters and 1.26 G FLOPs and outperforms 13 existing mainstream segmentation methods (Dice coefficient, mIoU) on multiple ultrasound datasets including placenta, gestational sac, thyroid nodules.
Outcomes and Implications
Because the model is lightweight and accurate, it is well-suited for clinical environments with limited computing resources (e.g., ultrasound machines in obstetrics). It may help improve automated segmentation of the placenta in ultrasound images, which is important because the shape and size of the placenta are closely related to fetal development in the second and third trimesters.