Related Experiment Video
Updated: May 24, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
SAM-MedUS: a foundational model for universal ultrasound image segmentation
Feng Tian1, Jintao Zhai2, Jinru Gong1
1Hunan Normal University, The School of Physics and Electronics, Changsha, China.
Journal of Medical Imaging (Bellingham, Wash.)
|March 3, 2025
Summary
This study introduces SAM-MedUS, a new model for segmenting diverse ultrasound images, outperforming existing methods. It enhances medical image analysis by improving segmentation accuracy across various anatomical sites.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate segmentation of ultrasound images is vital for medical diagnosis, monitoring, and research.
- Current segmentation methods are often limited to specific organs, tumors, or imaging devices.
- Existing deep learning models, like SAM-med2d, have limited application to ultrasound due to a scarcity of relevant training data.
Purpose of the Study:
- To develop a generic ultrasound image segmentation model, SAM-MedUS, capable of handling diverse ultrasound data.
- To create a comprehensive ultrasound dataset encompassing eight site categories for robust model training and validation.
- To enhance the model's ability to segment fuzzy boundaries and low-contrast regions characteristic of ultrasound imaging.
Main Methods:
- Proposed the SAM-MedUS model, integrating ConvNeXt V2 and CM blocks in the encoder for improved global context extraction.
- Utilized a newly curated, diverse ultrasound image dataset with eight site categories for training and testing.
- Incorporated a boundary loss function to specifically address challenges in segmenting fuzzy boundaries and low-contrast ultrasound images.
Main Results:
- SAM-MedUS demonstrated superior performance compared to recent methods across multiple ultrasound datasets.
- Achieved high segmentation accuracy on easier datasets (e.g., adult kidney: 87.93% IoU, 93.58% Dice).
- Showcased robust performance on complex datasets (e.g., infant vein: 62.31% IoU, 78.93% Dice).
Conclusions:
- The developed SAM-MedUS model offers uniform and accurate segmentation for a wide range of ultrasound images.
- The integration of advanced architectural components and a specialized loss function enhances global information extraction and boundary delineation.
- The model exhibits strong performance and excellent generalization capabilities, addressing limitations of previous ultrasound segmentation techniques.
