Deep neural-network based optimization for the design of a multi-element surface magnet for MRI applications

Area of Science:

• Computational physics
• Artificial intelligence
• Medical imaging

Background:

• Inverse problems are crucial in scientific and engineering fields.
• Traditional methods for solving inverse problems often require large datasets and regularization terms.
• Existing models for specialized applications like MRI magnet design can be computationally intensive and may suffer from overfitting.

Purpose of the Study:

• To introduce a novel encoder-analytic (EA) hybrid model for solving inverse problems.
• To demonstrate that the EA model can train itself using an analytical forward map, eliminating the need for dedicated training datasets and separate regularization.
• To showcase the model's application in designing multi-element surface magnets for low-field magnetic resonance imaging (MRI) and compare its performance against benchmark algorithms.

Main Methods:

• Development of a hybrid model combining a Convolutional Neural Network (CNN)-based encoder with an analytical forward map.
• Direct learning approach where the model trains itself from the connected forward map, which also serves as a regularizer.
• Customization of the model to find specific target solutions or solutions adhering to given heuristics.
• Application of the EA model to the inverse design of multi-element surface magnets for low-field MRI.

Main Results:

• The encoder-analytic (EA) hybrid model successfully solves inverse problems without requiring a dedicated training dataset or a separate regularization term.
• The forward map inherently acts as a regularizer, preventing the model from overfitting.
• The EA model demonstrated flexibility in finding specific solutions or heuristic-driven designs.
• In the application to MRI magnet design, the EA model significantly outperformed the benchmark genetic algorithm, achieving nearly 10 times better results.

Conclusions:

• The encoder-analytic (EA) hybrid model offers an efficient and effective approach to solving inverse problems.
• This data-independent and self-regularizing model presents a significant advancement over traditional methods.
• The EA model's successful application in MRI magnet design highlights its potential for optimizing complex engineering tasks and advancing medical imaging technology.