Randomized trial demonstrates improved segmentation accuracy in ultrasound images, suggesting a practical approach for real-time diagnosis of breast tumors.
Key Points
To improve semantic segmentation of malignant tumors in breast ultrasound images using a lightweight U-Net model.
Proposed a lightweight U-Net-based semantic segmentation algorithm with multi-scale supervision.
Implemented channel attention mechanisms for enhanced tumor feature extraction.
Performed structured pruning and quantized attention training to reduce model size and improve speed.
Achieved Dice coefficient of 0.818 for segmentation accuracy.