Efficient Breast Lesion Segmentation From Ultrasound Videos Across Multiple Source-Limited Platforms.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Medical video segmentation is crucial for clinical diagnosis and treatment, particularly for tracking breast lesions in ultrasound videos.
• Current methods struggle to balance segmentation accuracy with fast inference speeds, limiting real-time use in resource-limited settings.

Purpose of the Study:

• To introduce BaS, an efficient on-device breast lesion segmentation model designed for high performance and speed.
• To address the limitations of existing models in real-time medical video analysis.

Main Methods:

• Developed BaS, a breast lesion segmentation model integrating Stem module and BaSBlock for inter- and intra-frame analysis.
• Released two versions: BaS-S for enhanced segmentation and BaS-L for faster inference.

Main Results:

• BaS demonstrates superior segmentation efficiency and prediction accuracy compared to existing models on resource-constrained devices.
• The model effectively refines representations through advanced frame analysis techniques.

Conclusions:

• BaS significantly advances efficient medical video segmentation, offering a practical solution for real-time clinical applications.
• The model's adaptability makes it suitable for deployment across various medical platforms.