MoDL: Model-Based Deep Learning Architecture for Inverse Problems.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Model-based image reconstruction is crucial for various imaging modalities.
• Deep learning approaches, particularly convolutional neural networks (CNNs), have shown promise in improving reconstruction quality.
• Existing methods often struggle with complex forward models or require extensive training data.

Purpose of the Study:

• To introduce a novel model-based image reconstruction framework utilizing a CNN-based regularization prior.
• To develop a systematic approach for deriving deep architectures tailored for inverse problems with arbitrary structures.
• To enhance reconstruction performance while minimizing computational and data requirements.

Main Methods:

• A model-based image reconstruction framework incorporating a CNN as a regularization prior.
• End-to-end training with weight sharing across iterations for CNN customization to the forward model.
• Integration of numerical optimization blocks, such as the conjugate gradients algorithm, for enforcing data consistency within the network.

Main Results:

• The proposed framework requires smaller networks and less training data/time compared to direct inversion methods.
• CNN weights customized to the forward model yield superior performance over pre-trained denoisers.
• Decoupling network complexity from the number of iterations reduces overfitting and memory footprint.
• Using conjugate gradients for data consistency leads to faster convergence and improved performance, especially under memory constraints.

Conclusions:

• The developed CNN-based regularization prior offers an efficient and effective approach for model-based image reconstruction.
• This method demonstrates significant advantages in terms of data efficiency, performance, and computational resource utilization.
• The framework provides a flexible and powerful tool for tackling complex inverse problems in imaging.