Synopsis

Keywords: Neuro: Brain

Deep learning (DL)-based denoising and image enhancement techniques reduce scan time while improving SNR and maintaining spatial resolution. Combining DL-based denoising with other rapid imaging techniques including parallel imaging and compressed sensing further reduces scan time. DL-based denoising techniques may be particularly beneficial for potentially low SNR images and/or time-consuming sequences such as DWI with high b-value and large number of MPG directions, and it may improve image quality of quantitative maps. DL-based resolution enhancement such as super-resolution model is superior to conventional methods. To establish clinically useful DL-based denoising and image enhancement techniques, prospective multi-site studies are required.

Introduction

Developments in artificial intelligence (AI) have made a significant impact in clinical medicine. In particular, the development of deep learning in medical imaging is remarkable, and AI has been applied to MR images for various purposes including image quality improvement, segmentation, lesion detection and classification. In this section, clinical applications of deep learning for denoising and image enhancement in brain MR imaging will be discussed.

Applications of deep learning-based denoising and image enhancement

For MR image acquisition, there is a trade-off between scan time, signal-to-noise ratio (SNR), and tissue contrast. To overcome this issue, deep learning-based denoising and image enhancement techniques have been proposed. Those techniques reduce scan time while improving SNR and maintaining spatial resolution. Among various types of deep neural networks, convolutional neural networks (CNNs) are the most commonly used neural network architecture. We have developed a deep learning approach to reduce image noise of low SNR images using high SNR images as supervised images (1). Because this technique is applicable for multi-contrast images, standard protocol brain MR examination time can be accelerated. Clinical studies have shown the usefulness of deep learning-based denoising techniques improve image quality in conventional images, MRA and DTI (1-3). Combining deep leaning-based denoising techniques with other rapid imaging techniques including parallel imaging and compressed sensing further reduces scan time (4). Deep learning-based denoising techniques may be particularly beneficial for potentially low SNR images and/or time-consuming examinations such as diffusion-weighted images with high b-value and large number of MPG directions, arterial spin labeling and synthetic MR imaging. In those sequences, denoising may improve the image quality of quantitative maps as well as reduce scan time (5). Resolution enhancement is another approach for reducing scan time and improving image quality. Zero-padding to increase the matrix size of the final image has been widely applied. Also, deep learning approaches have been developed in image resolution enhancement. Among them, super-resolution is a promising technique. The performance of deep-learning super-resolution is superior to conventional resolution enhancement methods (6).

Clinical validation

Several multi-center studies showed the usefulness of deep-learning approaches for denoising and image enhancement in clinical brain MR imaging (7); however, most studies are limited in a small cohort at a single imaging site. Prospective multi-site studies are required to establish clinically useful deep learning-based denoising and image enhancement techniques.

Acknowledgements

I would like to thank Professor Toshinori Hirai and Dr. Hiroyuki Uetani, Department of Diagnostic Radiology, Kumamoto University, for useful discussions.