A multi-scale attention residual-based U-Net network for stroke electrical impedance tomography.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Neuroscience

Background:

• Electrical impedance tomography (EIT) is a non-invasive, radiation-free imaging method valuable for stroke diagnosis.
• EIT imaging faces challenges with low spatial resolution due to soft-field nonlinearity and ill-posed inverse problems.

Purpose of the Study:

• To develop an advanced deep learning network for improved stroke reconstruction in EIT.
• To address the limitations of low spatial resolution and artifacts in EIT-based stroke imaging.

Main Methods:

• A novel multi-scale convolutional attention residual-based U-Net (MARU-Net) was proposed, integrating residual and multi-scale attention modules into a U-Net architecture.
• The MARU-Net was applied to the EIT system for stroke imaging, enhancing feature extraction and information processing.
• Performance was evaluated against traditional convolutional neural networks and 1D convolutional neural networks.

Main Results:

• MARU-Net demonstrated superior performance with fewer artifacts and clearer reconstructed images compared to other methods.
• The network effectively reduced noisy artifacts, achieving an image correlation coefficient greater than 0.87 for noisy reconstructed images.
• Model physics experiments validated the practical applicability of the MARU-Net for EIT stroke imaging.

Conclusions:

• The MARU-Net network significantly improves the quality of EIT images for stroke reconstruction.
• This method offers a promising, high-resolution, and artifact-reduced imaging solution for stroke diagnosis using EIT.
• The study validates the effectiveness and practicality of MARU-Net in real-world EIT applications.