U-Net-based approach for automatic lung segmentation in electrical impedance tomography.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Electrical impedance tomography (EIT) is a non-invasive bedside tool for monitoring lung ventilation.
• Current methods struggle with identifying lung regions in EIT images, especially in patients with poor ventilation, due to low spatial resolution.
• Accurate lung segmentation is crucial for assessing ventilation-dependent parameters for therapy.

Purpose of the Study:

• To develop an automatic, robust, and rapid lung segmentation method for EIT images.
• To define regions-of-interest (ROIs) for improved analysis of ventilation distribution.
• To overcome limitations of existing methods in identifying lung areas within EIT data.

Main Methods:

• A U-Net-based deep learning model was employed as a postprocessor for EIT images.
• The model automatically segments lung ROIs and refines conductivity distribution without prior information.
• A finite element method (FEM) phantom was used for experimental validation.

Main Results:

• The U-Net model achieved automatic segmentation of lung ROIs in EIT images.
• Simulations using an FEM phantom demonstrated high accuracy with a Dice similarity coefficient (DSC) >0.99 and mean absolute error (MAE) of 0.0065.
• The method provides distinguishable lung ROIs for parameter extraction.

Conclusions:

• Deep learning enables convenient and automatic segmentation of lung ROIs from EIT images.
• This facilitates the extraction and analysis of regional lung ventilation parameters.
• Further validation on human datasets is needed to refine the model for clinical application.