Two-stage deep learning network-based few-view image reconstruction for parallel-beam projection tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Projection tomography (PT) is vital for volumetric imaging with isotropic resolution.
• Few-view PT is crucial for reducing scan time, radiation dose, and simplifying sample fixation.
• Current few-view PT reconstruction techniques are limited, especially for in vivo applications.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for parallel-beam PT reconstruction from few-view projection data.
• To improve the accuracy and feasibility of few-view PT reconstruction, particularly for in vivo imaging.

Main Methods:

• A 2-stage deep learning network (TSDLN) framework was proposed, comprising a reconstruction network (R-net) and a correction network (C-net).
• The R-net, a generative adversarial network (GAN), completes image information from direct back-projection of sparse signals.
• The C-net, a U-net array, denoises the reconstruction for high-quality output.

Main Results:

• The TSDLN framework demonstrated superior reconstruction performance compared to traditional methods on the DeepLesion dataset, especially with sparse-view data.
• High-quality images were reconstructed from as few as two projections, with structural similarities greater than 0.8.
• The framework showed migration capabilities, successfully reconstructing images from optical PT data trained on CT data, preserving contours and distribution information with only 5 projections.

Conclusions:

• The TSDLN-based framework exhibits robust reconstruction capabilities for few-view projection images.
• This deep learning approach holds significant potential for advancing in vivo projection tomography applications.