An evolutionary neural architecture search for magnetic resonance image reconstructions

Deep learning has emerged as a promising approach for magnetic resonance imaging (MRI) reconstruction from undersampled data. However, manually designing an optimized neural network architecture is time-consuming and may not yield the most efficient or accurate results. This study proposes an automatic method to design a convolutional neural network (CNN) model optimized for MRI reconstruction accuracy.

We introduce a genetic algorithm framework that optimizes CNN architectures by tuning fractional representations of architectural components, rather than conventional hyperparameters. This encoding enables flexible exploration of the design space, increasing the likelihood of identifying the optimal architecture more efficiently. By evolving these parameters through successive generations guided by model performance, the method efficiently identifies optimal configurations tailored for MRI reconstruction, eliminating the need for manual architecture engineering.

The method was evaluated on the fastMRI dataset provided by New York University. Results show that the proposed genetic algorithm (GA)-based automatically designed network achieved higher accuracy than manually designed CNNs. On brain MRI reconstructions, the proposed method outperformed existing architectures. Application to a knee dataset further demonstrated the algorithm's ability to discover even more effective models, highlighting its adaptability and generalization across anatomy types.

The study confirms that automated CNN design using GA can significantly improve MRI reconstruction quality while eliminating the need for manual tuning. This approach offers an efficient and scalable solution for developing high-performance deep learning models for medical imaging applications.