SAM-MedUS

Ultrasound imaging is one of the most widely used clinical modalities, but automated segmentation of ultrasound images is notoriously difficult: images suffer from speckle noise, low contrast, blurred boundaries, and large appearance differences between anatomical sites. Most deep-learning segmentation models are trained for a single organ or task, which limits their generality and forces clinicians to maintain many narrow models. SAM-MedUS is a foundational model that addresses this fragmentation by targeting universal ultrasound segmentation across diverse anatomies in a single network.

Developed by Feng Tian and colleagues at Hunan Normal University, with collaboration from Hunan University (Changsha, China), SAM-MedUS builds on the Segment Anything Model (SAM) paradigm of promptable, transformer-based segmentation. Rather than fine-tuning SAM for one organ, the authors assemble a broad multi-anatomy ultrasound corpus and augment SAM's encoder to better capture the global context and fuzzy boundaries that characterize sonographic data. The work was published in the Journal of Medical Imaging in February 2025.

The model sits alongside other SAM adaptations for ultrasound (such as SAMUS and UltraSam), but distinguishes itself through its explicit emphasis on cross-anatomy generality and boundary-aware optimization, aiming for a single backbone that transfers across thyroid, breast, kidney, fetal, vascular, and other ultrasound targets.

#Key Features

• Universal multi-anatomy scope: A single model is trained and evaluated across eight anatomical categories—thyroid, breast, liver, kidney, ovarian tumor, fetal head, brachial nerve, and infant vein—rather than one organ at a time.
• SAM-based promptable architecture: Retains SAM's image encoder, prompt encoder (points and bounding boxes), and iterative mask decoder, enabling interactive and automatic segmentation workflows.
• ConvNeXt V2 and CM blocks: The encoder integrates a ConvNeXt V2 auxiliary branch and combined depthwise/pointwise (CM) convolution blocks to strengthen global context extraction on noisy ultrasound images.
• Boundary-aware loss: A dedicated boundary loss term emphasizes edge accuracy, improving delineation of the fuzzy, low-contrast boundaries that are common in sonography.

#Technical Details

SAM-MedUS uses a Vision Transformer (ViT)-based image encoder augmented with a ConvNeXt V2 branch and CM convolution blocks, paired with SAM's standard prompt encoder and mask decoder. Training data comprised roughly 15 open-source ultrasound collections grouped into eight anatomical categories (thyroid datasets such as TG3K, TN3K, and DDTI dominate at about 38.7% of images, with breast datasets contributing the largest number of constituent collections). All images were standardized to 256×256 resolution and split 80/20 for training and testing. The model was trained in PyTorch on an NVIDIA RTX 3090 using the Adam optimizer with an initial learning rate of 1e-4. Across all anatomies, SAM-MedUS reached an average Dice of about 88.0% and IoU of about 78.9%, with kidney (93.6% Dice) and fetal head (96.0% Dice) among the strongest, and infant vein the most challenging (78.9% Dice). It outperformed comparison methods including U-Net, SegNet, DANet, CMU-NeXt, and VM-UNet on multiple datasets.

#Applications

SAM-MedUS is aimed at clinical and research ultrasound workflows where a single, general segmentation tool is preferable to many organ-specific models. Potential uses include delineating thyroid nodules, breast lesions, kidney structures, ovarian tumors, fetal head biometry, and vascular targets for measurement, diagnosis support, and downstream quantification. Its promptable design lets sonographers and radiologists guide segmentation interactively with clicks or boxes, while its multi-anatomy training makes it a candidate backbone for building broader ultrasound analysis pipelines without retraining a new model for each site.

#Impact

By demonstrating that a single SAM-derived backbone can segment eight distinct anatomical regions with competitive accuracy, SAM-MedUS contributes to the growing effort to bring foundation-model generality to medical ultrasound, a modality historically served by fragmented, task-specific networks. Its combination of multi-dataset aggregation, ConvNeXt V2 encoder enhancements, and boundary-aware loss offers a reusable recipe for cross-anatomy sonographic segmentation. A key limitation is openness: at publication the authors noted that code would be released after the paper, and pretrained weights and an explicit open-source license were not yet posted to the project repository, which currently constrains reproducibility and direct adoption.