An EIT image reconstruction method based on DenseNet with multi-scale convolution.

Area of Science:

• Biomedical Imaging
• Medical Physics
• Computational Imaging

Background:

• Electrical impedance tomography (EIT) non-invasively images electrical conductivity distributions.
• EIT image reconstruction is challenging due to its ill-posed and nonlinear nature, leading to lower image quality.
• Existing methods often struggle with artifacts and edge blurring.

Purpose of the Study:

• To propose an advanced EIT image reconstruction method, MS-DenseNet, to overcome the limitations of conventional techniques.
• To improve the quality and accuracy of EIT images by addressing artifacts and blurring.
• To enhance the generalization and fitting abilities of deep learning models for EIT.

Main Methods:

• Developed MS-DenseNet, incorporating parallel multi-scale convolutional dense blocks to enhance network generalization.
• Implemented a hybrid pooling structure in the connection layer to minimize information loss during pooling.
• Utilized a two-stage learning rate reduction strategy to optimize network fitting.
• Trained and validated the model using both simulated and real-world EIT measurement data.

Main Results:

• MS-DenseNet significantly improved EIT image quality compared to conventional DenseNet and Gauss-Newton methods.
• The proposed method demonstrated superior performance in reducing artifacts and edge blurring.
• Quantitative evaluation using RMSE, SSIM, MAE, and ICC showed marked improvements in image metrics.

Conclusions:

• MS-DenseNet offers a robust and effective solution for enhancing EIT image reconstruction.
• The integration of multi-scale convolutions and hybrid pooling contributes to superior image quality and accuracy.
• This method holds promise for advancing non-invasive imaging applications in various fields.