MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Super-Resolution 1H Magnetic Resonance Spectroscopic Imaging Utilizing Deep Learning

Z Iqbal*, D Nguyen , S Jiang , Medical Artificial Intelligence and Automation Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, USA 75390

Presentations

(Thursday, 8/2/2018) 7:30 AM - 9:30 AM

Room: Karl Dean Ballroom B1

Purpose: ¹H magnetic resonance spectroscopic imaging (SI) provides valuable biochemical information from the human brain but suffers from low resolution. In fact, it is impossible to obtain high-resolution spectroscopic images (HRSI) in vivo using vendor provided protocols. We hypothesize that super-resolution SI can be reconstructed using deep learning by incorporating anatomical information from a T1-weighted image (T1w) and metabolic information from a low-resolution spectroscopic image (LRSI).

Methods: We designed an SI generator that produces realistic HRSI and LRSI from an input T1w by leveraging known biological relationships. A novel densely connected UNet (D-UNet) was trained using the dataset synthesized from the SI generator, and this reconstruction method was validated using both simulated and in vivo results. We performed error calculations using mean squared error (MSE).

Results: The D-UNet successfully reconstructed HRSI (128x128resolution) from rescaled T1w (128x128resolution) and LRSI (24x24resolution) images (MSE<0.01 for training). The SI generator was invaluable for creating a large and diverse dataset for training (102,000 images), which would otherwise be impossible to obtain. The MSE for the reconstruction method (MSE=1.316) was low when compared to standard upscaling techniques such as zero-filling (MSE=1.652) and bicubic interpolation (MSE=3.129) for 169 testing data. The D-UNet method reconstructed entire ¹H spectra containing important brain metabolites accurately (MSE<0.15 for 512 spectral points). Finally, this method accurately reconstructed in vivo HRSI data (MSE<0.1).

Conclusion: We present a novel method capable of producing super-resolution SI using deep learning. The super-resolution method presented here has many potential applications. This method can accelerate current SI protocols, improve the quality of metabolite maps acquired from SI protocols, and improve analysis of metabolic images. This methodology may be applicable to other super-resolution applications if a relationship between the low-resolution image and the anatomy is well known.

Keywords

Spectroscopic Imaging, Brain, Resolution

Taxonomy

IM- MRI : Spectroscopy

Contact Email