Training Ultrasound Image Classification Deep-Learning Algorithms for Pneumothorax Detection Using a Synthetic Tissue...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound (US) is vital in emergency and military medicine due to its portability.
• Interpreting US images requires expertise, limiting remote use for conditions like pneumothorax (PTX).
• Artificial intelligence (AI) can automate US image analysis, but requires large datasets and validation.

Purpose of the Study:

• To develop a dynamic synthetic tissue phantom for training AI algorithms to detect pneumothorax (PTX).
• To evaluate the phantom's efficacy in mimicking PTX scenarios for AI model training.
• To demonstrate the utility of the phantom for AI applications in remote diagnostics.

Main Methods:

• A synthetic gelatin phantom was created using a 3D-printed rib mold and a lung-mimicking component.
• The phantom was used to acquire images simulating both PTX-negative and PTX-positive conditions.
• A deep learning image classification algorithm, previously used for shrapnel detection, was trained and tested on phantom images.

Main Results:

• Images acquired from the synthetic phantom were comparable to those from swine models in both PTX-negative and PTX-positive scenarios.
• The AI algorithm accurately predicted PTX in swine images after being trained exclusively on phantom-generated image sets.
• This demonstrates the phantom's capability to serve as a viable training dataset for AI.

Conclusions:

• The developed dynamic synthetic tissue phantom is a valuable tool for training AI algorithms for PTX detection.
• This approach reduces the need for extensive animal models or clinical testing in early AI development for ultrasound diagnostics.
• The phantom facilitates the advancement of AI-powered remote medical diagnostic tools.