An Efficient and Accurate Super-Resolution Approach to Low-Field MRI via U-Net Architecture With Logarithmic Loss and L2 Regularization

Overview

The medical field is becoming increasingly reliant on medical imaging for diagnostic procedures. Although traditional high-field MRI scanners produce high-quality images at high SNR, they are inefficient for emergency neuroimaging tasks and quick analysis of the patients’ conditions. Low field (LF) MRI scanners are portable, cost-effective, and safer tools for non-invasive diagnostic imaging. Such scanners can be wheeled to the patient’s bedside due to the low strength of the scanners (64 mT-brought down to 1/1000 of the magnetic field strength of traditional scanners). Here, I was able to generate high-resolution images from representative low SNR down-sampled scans via a machine learning approach. Traditional machine learning super-resolution (SR) reconstruction methods for medical imaging that utilize general architectures like static convolutional neural networks (SCNNs) or recurrent neural networks (ReCNNs) are unable to reconstruct precise anatomical figures. Hence, I proposed an SR U-Net to produce high-resolution images more accurately and efficiently. Unlike SCNNs and ReCNNs, U-Nets concave layers of the same shape to preserve feature upscaling which produces more anatomically correct results. My aim in this study was to accurately downsample high-resolution MRI scans to reflect image quality attainable at low field and utilize a machine learning reconstruction approach to increase their resolution. Here, I evaluated this novel approach on T1-weighted downsampled scans and actual LF MRI scans to demonstrate the efficiency of my network. Additionally, I incorporated a logarithmic loss component and L2 regularization to make the U-Net more efficient in learning and to prevent over smoothing.

Awards (3)

Second Award of $2,000 $2,000
National Taiwan Science Education Center: Taiwan International Science Fair Special Award is a trip to participate in the Taiwan International Science Fair
Drexel University: Full tuition scholarship

Competition history

ISEF 2022 Biomedical Engineering · Entry ENBM034 · Atlanta, Georgia, United States

Resources

ISEF abstract page

Related projects

ISEF · 2023

Low-Field MRI Super-Resolution With Densely Engineered U-Net++ Network: Improving Resolution and Contrast of Portable MRI With Nested U-Net Architecture, Enabling Low-Cost Bedside Neurological Imaging on a Global Scale