Dynamic Imaging Using Deep Bi-Linear Unsupervised Representation (DEBLUR).

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging
• Artificial Intelligence

Background:

• Bilinear models are effective for dynamic MRI reconstruction, but often rely on sparsity and energy compaction priors.
• Existing deep learning methods can be memory-intensive due to storing all time frames.

Purpose of the Study:

• To develop a novel bilinear model for dynamic MRI using convolutional neural networks (CNNs) to generate factor matrices.
• To enable memory-efficient and fast reconstruction of dynamic MRI data, particularly for free-breathing cardiac applications.

Main Methods:

• A novel bilinear model where factor matrices are generated by CNNs, with parameters learned from undersampled data.
• CNN parameters are initialized using existing factor methods and regularized with sparsity to prevent overfitting.
• The method was tested on free-breathing, ungated cardiac cine MRI data acquired with a navigated golden-angle radial sequence.

Main Results:

• The proposed method demonstrates reduced spatial blurring compared to classical bilinear methods.
• It also outperforms a recent unsupervised deep learning approach in terms of spatial blurring.
• The approach is memory-efficient, requiring storage only of factors or compressed representations.

Conclusions:

• The novel CNN-based bilinear model offers an efficient and effective solution for dynamic MRI reconstruction.
• This method is suitable for large-scale dynamic applications like free-breathing cardiac MRI.
• It provides improved image quality with reduced spatial blurring.