Research on Lung Ultrasound Image Classification Based on Compressed Sensing

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Emergency Medicine

Background:

• Pneumothorax diagnosis is critical in emergency settings but conventional imaging is impractical.
• Lung ultrasound is portable but suffers from image quality issues.
• Efficient and accurate diagnostic tools are needed for field use.

Purpose of the Study:

• To develop and evaluate a lung ultrasound image recognition and classification method for pneumothorax.
• To improve diagnostic efficiency and accuracy in resource-limited environments.
• To adapt AI models for portable, embedded diagnostic devices.

Main Methods:

• Collected lung ultrasound M-mode images from a simulation model.
• Applied compressed sensing for data compression and feature extraction.
• Designed and trained a BP neural network for image classification.
• Compared performance against the VGG16 network.

Main Results:

• The proposed compressed sensing and BP neural network method achieved higher recognition accuracy.
• The new method demonstrated significantly fewer parameters compared to VGG16.
• The approach is suitable for deployment on embedded systems.

Conclusions:

• Compressed sensing combined with BP neural networks offers an effective solution for lung ultrasound image analysis.
• This method improves diagnostic accuracy and efficiency for pneumothorax detection in emergency scenarios.
• The reduced model complexity makes it ideal for portable, embedded diagnostic tools.