Deep supervised dictionary learning by algorithm unrolling-Application to fast 2D dynamic MR image reconstruction.

Area of Science:

• Medical Imaging
• Machine Learning
• Signal Processing

Background:

• Unrolled neural networks (NNs) are widely used for image reconstruction, enabling physics-informed learning of regularization but often acting as black boxes.
• Dictionary learning (DL) is a regularization technique that learns transforms for sparse signal approximation, typically pretrained or jointly trained without considering the reconstruction physics.

Purpose of the Study:

• To develop a dictionary learning (DL) algorithm based on unrolled neural networks (NNs) that integrates the reconstruction problem and physical imaging model.
• To overcome limitations of existing DL algorithms by incorporating physics-informed learning within an unrolled NN framework.

Main Methods:

• Constructed an unrolled NN corresponding to a DL-based reconstruction algorithm, optimizing dictionary atoms via back-propagation.
• Employed a novel 2D dictionary in the spatio-temporal domain for accelerated cardiac cine MR image reconstruction.
• Evaluated performance against state-of-the-art deep iterative CNNs on dynamic MRI datasets.

Main Results:

• The physics-informed DL approach achieved significantly more accurate reconstructions (up to 4.90 dB PSNR, 5% SSIM improvement) compared to decoupled pretraining.
• The spatio-temporal 2D dictionary enhanced accuracy (up to 1.10 dB PSNR, 4% SSIM) by preserving details while reducing artifacts and noise.
• The proposed NN-based DL method demonstrated superior stability and interpretability compared to CNNs, yielding comparable results on a second dataset.

Conclusions:

• The physics-informed NN serves as a training algorithm for interpretable, data-driven dictionary learning regularization.
• This approach links learned dictionaries to specific data and the reconstruction method defined by the NN, enhancing transparency and applicability.