A Green Prospective for Learned Post-Processing in Sparse-View Tomographic Reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Imaging

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise for medical imaging reconstruction from limited data.
• Existing deep CNN architectures are often too deep, leading to overfitting and limiting clinical feasibility.
• Filtered backprojection (FBP) is a common but often insufficient reconstruction algorithm for sparse data.

Purpose of the Study:

• To propose a shallow neural network for efficient learned post-processing in medical image reconstruction.
• To address the limitations of deep CNNs in terms of computational cost and clinical applicability.
• To improve the quality and efficiency of tomographic reconstructions from subsampled measurements.

Main Methods:

• Developed a shallow neural network for learned post-processing.
• Applied the network to images reconstructed by the filtered backprojection algorithm.
• Compared the proposed network's performance against a deep ResUNet architecture.
• Evaluated reconstructions from sparse-view computed tomography protocols.

Main Results:

• The proposed shallow network achieved comparable or higher image quality than the deep ResUNet.
• The shallow network computed reconstructions in approximately one-fourth of the time.
• The shallow network demonstrated increased robustness in tomographic reconstructions.
• The approach aligns with Green AI principles by reducing computational demands.

Conclusions:

• A shallow neural network offers an efficient and effective alternative for medical image reconstruction.
• Learned post-processing with shallow networks is a viable strategy for improving sparse-view CT reconstruction.
• This Green AI-inspired approach enhances clinical feasibility by reducing computational load and time.