Computationally efficient deep neural network for computed tomography image reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep neural networks show promise for medical image reconstruction in low-dose and undersampled scenarios.
• Training these networks, especially for 3D CT, is computationally intensive, requiring substantial memory and time.
• Current hardware limitations hinder the practical application of deep learning for high-resolution 3D CT reconstruction.

Purpose of the Study:

• To develop a computationally efficient neural network for CT image reconstruction.
• To reduce the memory and time demands of training CT reconstruction networks.
• To enable practical 3D CT reconstruction on mainstream hardware without compromising image quality.

Main Methods:

• Unrolled proximal gradient descent algorithm integrated with convolutional neural networks (CNNs).
• Greedy training on image patches in an iteration-by-iteration manner.
• Utilized deep UNet architecture with separable quadratic surrogate and ordered subsets for enhanced data fidelity and to mitigate local minima.

Main Results:

• Achieved comparable image quality to state-of-the-art methods for 2D sparse-view and limited-angle CT reconstruction.
• Demonstrated significant reductions in training requirements: 2 GB GPU memory and 0.45 s/iteration minimum.
• Outperformed traditional iterative methods (Total Variation, Dictionary Learning) in both 2D and 3D CT reconstruction.

Conclusions:

• A computationally efficient neural network for CT image reconstruction was developed.
• The proposed method offers comparable image quality to existing deep learning approaches with drastically reduced training resource needs.
• The method is suitable for 3D reconstruction tasks like cone-beam CT and tomosynthesis on standard GPUs.