Adaptive convolutional neural networks for accelerating magnetic resonance imaging via k-space data interpolation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning accelerates magnetic resonance imaging (MRI) by reconstructing images from undersampled k-space data.
• Existing methods often overlook k-space spatial frequency properties and adjacent slice information, limiting reconstruction accuracy.
• Convolutional Neural Networks (CNNs) are common but can be ineffective without considering k-space characteristics.

Purpose of the Study:

• To develop an advanced deep learning algorithm for improved k-space data interpolation in MRI.
• To address limitations of current methods by incorporating spatial frequency properties and multi-slice information.
• To enhance the accuracy and efficiency of image reconstruction from undersampled k-space data.

Main Methods:

• Developed adaptive convolutional neural networks for k-space data interpolation (ACNN-k-Space) using a residual Encoder-Decoder architecture.
• Integrated spatially contiguous slices as multi-channel input and utilized multi-coil data.
• Incorporated self-attention layers to adaptively focus on k-space data across spatial frequencies and channels.

Main Results:

• Evaluated ACNN-k-Space on two public datasets, comparing it against state-of-the-art methods.
• Demonstrated effective image reconstruction from undersampled k-space data.
• Achieved significantly superior image reconstruction performance compared to existing techniques.

Conclusions:

• ACNN-k-Space effectively reconstructs images from undersampled k-space data, outperforming current state-of-the-art methods.
• The integration of spatial frequency awareness and multi-slice information enhances deep learning-based MRI reconstruction.
• The developed algorithm offers a promising advancement for fast and accurate MRI image reconstruction.