An adaptive attention U-network for recognizing ultrasound images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anesthesiology

Background:

• Traditional intraspinal anesthesia relies on surface landmarks, leading to low accuracy and procedural complexity.
• Accurate identification of spinal anatomical structures is crucial for safe and effective anesthesia delivery.

Purpose of the Study:

• To develop an adaptive attention U-network for enhanced segmentation of spinal structures in ultrasound images.
• To improve the accuracy and efficiency of spinal anatomical structure identification for intraspinal anesthesia.

Main Methods:

• Collected 1000 annotated ultrasound images from 80 pregnant women to create a spine ultrasound image dataset.
• Developed an adaptive attention U-network utilizing multidepth convolution kernels and adaptive local channel attention modules.
• Incorporated global attention gate module and multiscale adaptive dynamic modulation for feature extraction and image enhancement.

Main Results:

• Achieved a mean Dice Similarity Coefficient of 0.905 on the spine ultrasound image dataset.
• Demonstrated superior generalization with a Dice Similarity Coefficient of 0.857 for benign tumor segmentation in a breast ultrasound dataset.
• Showcased consistent segmentation stability across all tested structures.

Conclusions:

• The adaptive attention U-network significantly improves segmentation accuracy for spinal anatomical structures in ultrasound images.
• The proposed model offers superior precision compared to existing medical image segmentation methods.
• This advancement has the potential to enhance the safety and efficacy of intraspinal anesthesia.