Deep Convolutional Neural Network for Inverse Problems in Imaging.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Ill-posed inverse problems are common in medical imaging, often solved with computationally intensive iterative algorithms.
• Existing methods face challenges with high computational costs and hyperparameter tuning.
• Unrolled iterative methods resemble Convolutional Neural Networks (CNNs) when the forward model's normal operator is a convolution.

Purpose of the Study:

• To propose a novel CNN-based algorithm for solving ill-posed inverse problems, specifically normal-convolutional inverse problems.
• To leverage the structure of unrolled iterative methods for CNN development.
• To improve computational efficiency and image quality in sparse-view reconstruction.

Main Methods:

• Developed a direct inversion followed by a CNN approach.
• The CNN incorporates multiresolution decomposition and residual learning.
• Applied the method to sparse-view parallel beam X-ray computed tomography reconstruction.

Main Results:

• The proposed CNN algorithm effectively removes artifacts from direct inversion while preserving image structure.
• Achieved high-quality sparse-view reconstruction (down to 50 views) in both synthetic and real X-ray computed tomography data.
• Outperformed total variation-regularized iterative reconstruction on realistic phantoms.
• Reconstructed a 512x512 image in under a second on a GPU.

Conclusions:

• The proposed CNN-based algorithm offers an efficient and effective solution for normal-convolutional ill-posed inverse problems.
• Demonstrated superior performance compared to traditional iterative methods in sparse-view X-ray computed tomography.
• Highlights the potential of deep learning for accelerating and enhancing medical image reconstruction.