A neural network with encoded visible edge prior for limited-angle computed tomography reconstruction

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning in Medical Diagnostics

Background:

• Limited-angle computed tomography (CT) is crucial for medical and industrial applications but is highly ill-posed, leading to artifacts in reconstructions.
• Conventional methods often rely on strong prior assumptions (e.g., piecewise constant images) that may not hold true, limiting their effectiveness.
• Convolutional Neural Networks (CNNs) show promise for modeling complex data relationships in medical imaging, but robustness remains a challenge.

Purpose of the Study:

• To propose a novel deep-learning approach, the alternating edge-preserving diffusion and smoothing neural network (AEDSNN), for improved limited-angle CT reconstruction.
• To leverage the theory of visible and invisible boundaries by incorporating visible edge information as a structural prior within the AEDSNN.
• To generalize and enhance the alternating edge-preserving diffusion and smoothing (AEDS) method by replacing its regularization terms with CNNs, thereby relaxing the piecewise constant assumption.

Main Methods:

• The AEDSNN is developed by unrolling the AEDS algorithm, with each block representing an iteration comprising data matching, x-direction, and y-direction regularization layers.
• The data matching layer utilizes the ordered-subset simultaneous algebraic reconstruction technique (OS-SART), while regularization layers are implemented using CNNs for intelligent prior encoding.
• Attention mechanisms and pooling layers are integrated into AEDSNN to strengthen visible edge priors and facilitate edge-preserving diffusion.

Main Results:

• AEDSNN demonstrated superior performance compared to popular limited-angle reconstruction algorithms on medical datasets, effectively handling piecewise smooth images with non-sharp edges.
• Experiments on printed circuit board (PCB) datasets confirmed AEDSNN's ability to better encode and utilize visible edge priors, yielding consistently improved reconstructions.
• The proposed method successfully overcomes the limitations of the piecewise constant assumption inherent in many conventional reconstruction algorithms.

Conclusions:

• The proposed deep-learning based AEDSNN offers a significant advancement in limited-angle CT reconstruction, outperforming existing methods.
• AEDSNN is free from parameter-tuning, operates rapidly (3-5 iterations vs. hundreds/thousands), and its learned regularizer demonstrates broad applicability to piecewise smooth images.
• This approach provides a robust and efficient solution for limited-angle reconstruction challenges in both medical and industrial imaging.