TranIU-Net: Indicative Electrical Tomography Imaging Based on Implicit Unrolling Transformer.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep unrolling networks combine data-driven and model-driven approaches for image reconstruction.
• Existing methods face challenges like inadequate iterations, limited receptive fields, and isolation from physical systems.

Purpose of the Study:

• To propose TranIU-Net, a novel implicit unrolling Transformer architecture for enhanced image reconstruction.
• To address limitations of current unrolling networks by integrating local and non-local dependencies and guiding system design.

Main Methods:

• TranIU-Net unrolls the proximal gradient algorithm into a trainable network with structural interpretability.
• An embedded Transformer module captures multi-scale information with hybrid receptive fields and a significance estimator.
• An implicit mapping guarantees convergence at unlimited depth with constant memory cost.

Main Results:

• TranIU-Net demonstrates superior performance over state-of-the-art methods in electrical tomography reconstruction.
• The architecture achieves improved quantitative and qualitative reconstruction results across various scenarios.
• The method effectively bridges the gap between reconstruction algorithms and imaging systems.

Conclusions:

• TranIU-Net offers a robust and efficient solution for image reconstruction tasks.
• The proposed architecture advances the field by integrating advanced deep learning techniques with model-driven principles.
• This work facilitates better reconstruction quality by considering intrinsic image correlations and system design.