Answers are grounded in this paper and cite their sources.
Demonstrates improved segmentation accuracy in ultrasound images, suggesting a practical approach for real-time diagnosis of breast tumors.
To address the challenges of high noise, blurred edges in breast ultrasound images, and the large parameter size and slow inference speed of traditional U-Net models, which fail to meet the demands of real-time clinical auxiliary diagnosis, a lightweight U-Net-based semantic segmentation method for malignant tumors in breast ultrasound images with multi-scale supervision is proposed. This algorithm adopts the classical U-Net architecture as its foundation, introduces multi-scale supervision strategies and channel attention mechanisms to enhance tumor feature extraction capabilities, while simultaneously employing structured pruning and quantized attention training (QAT) to achieve extreme model lightweighting. Experimental results demonstrate that the lightweight model maintains segmentation accuracy (Dice coefficient reaching 0.818) while reducing parameter size by approximately 76.8% and improving inference speed by about 77.7%. This approach provides efficient technical support for intelligent auxiliary diagnosis of malignant breast tumors, making it suitable for low-power, real-time diagnostic scenarios in primary healthcare institutions.
No takes yet. Share an insight, caveat, or question.
Li Yusen (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: