Echo Flow-Induced Temporal Correlation Learning for Ultrasound Video Object Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Ultrasound video object segmentation is crucial for delineating anatomical structures and injuries.
• Existing methods struggle with inter-frame motion and low-contrast scenarios common in ultrasound.
• While SAM2 advanced video segmentation, its performance on ultrasound data is limited due to its natural video bias.

Purpose of the Study:

• To develop a novel method, EchoSAM2, for accurate object segmentation in ultrasound videos.
• To address limitations in capturing temporal dynamics and ultrasound-specific characteristics.
• To improve the clinical value of ultrasound diagnostics through enhanced segmentation.

Main Methods:

• Proposed Echo Flow to model inter-frame motion trends and suppress non-object interference.
• Introduced an Echo Modulation Block (EMB) to integrate Echo Flow into the SAM2 framework.
• Developed a Gaussian Adapter for optimizing SAM2 fine-tuning on ultrasound images.

Main Results:

• EchoSAM2 achieved state-of-the-art results on three ultrasound video datasets.
• Achieved a Dice score of 85.49% on the EUDP dataset, surpassing the second-best method by 3.19%.
• Demonstrated strong generalization capabilities on unseen test sets from different datasets.

Conclusions:

• The Echo Flow and supporting modules significantly improve segmentation accuracy and model generalizability.
• Enhanced ultrasound video segmentation leads to improved diagnostic accuracy and clinical value.