Projection-to-image transform frame: a lightweight block reconstruction network for computed tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Learning-based computed tomography (CT) reconstruction methods face challenges with high memory requirements.
• Existing neural networks for tomographic reconstruction are often computationally intensive.

Purpose of the Study:

• To introduce a novel lightweight block reconstruction network (LBRN) for computed tomography (CT) reconstruction.
• To overcome the memory space limitations of current deep learning approaches in CT.

Main Methods:

• The LBRN unrolls the filter back-projection (FBP) algorithm into a deep neural network.
• It features two modules: one for filtering and one for block back-projection, decoupling the Radon transform.
• The network is trained end-to-end using raw projection data without requiring initial images.

Main Results:

• The LBRN significantly simplifies the network structure and reduces the number of parameters.
• Experimental results demonstrate the network's effectiveness across various CT reconstruction scenarios.
• The approach shows outstanding advantages in full angle, low-dose CT, region of interest, and metal artifact reduction.

Conclusions:

• The proposed lightweight block reconstruction network (LBRN) is an effective deep learning solution for CT reconstruction.
• LBRN offers a memory-efficient and parameter-reduced alternative to existing methods.
• This approach shows broad applicability and significant advantages in diverse CT reconstruction tasks.