Image reconstruction by domain-transform manifold learning.

Area of Science:

• * Computational imaging and machine learning.
• * Development of novel algorithms for scientific imaging.

Background:

• * Image reconstruction is critical across diverse scientific fields, from medical imaging to astronomy.
• * Conventional methods often rely on complex, ad hoc signal processing chains requiring expert tuning.
• * Challenges include unknown inverse transforms and sensor imperfections like noise.

Purpose of the Study:

• * To introduce a unified, data-driven framework for image reconstruction called AUTOMAP.
• * To demonstrate AUTOMAP's flexibility and effectiveness using deep learning.
• * To improve reconstruction performance and robustness against noise and artifacts.

Main Methods:

• * Implemented AUTOMAP using a deep neural network architecture.
• * Trained the network on a corpus of data to learn sensor-to-image domain mappings.
• * Utilized manifold learning to achieve sparse representations of transforms.

Main Results:

• * AUTOMAP successfully learned reconstruction transforms for various magnetic resonance imaging strategies with consistent hyperparameters.
• * Demonstrated superior noise immunity and reduced reconstruction artifacts compared to traditional methods.
• * Showcased manifold learning's role in creating sparse, low-dimensional data representations.

Conclusions:

• * AUTOMAP provides a flexible and powerful data-driven alternative to conventional image reconstruction.
• * Learned reconstruction approaches, like AUTOMAP, can enhance existing imaging techniques.
• * This framework is expected to accelerate the development of new imaging acquisition strategies.