MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Structural and Functional Magnetic Resonance Imaging (MRI) Super-Resolution Using Deep Convolutional Neural Network

H Liu*, University of Florida, Gainesville, FL

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: MRI is proved to be a safe, non-invasive and effective way of evaluating brain function. It does not use radiation as x-ray or CT, and is far more objective compared to the traditional questionnaire methods of psychological evaluation. However, the trade-off between motion artifact, which is mostly caused by long acquisition time, and image quality has been a problem for researchers.
Methods: study tackles the MRI optimization problem by integrating image details learned from structural MR images and generating high-quality fMRI as outputs. We use a deep learning-based method combined with transfer learning in order to maintain the functional MR image quality and reduce the subject motion artifacts in scenarios where longer acquisition time is not preferable. The dataset were selected from a rodent imaging project where high-field magnetic scanner was used. The protocol was approved by IACUC, where functional and structural MRI in low and high-resolution for each rodent brain were acquired. Our method is evaluated on 4875 images collected from 16 subjects, including 195 T2 structural MRI and 4680 fMRI slices. We down-sample the high-resolution images into a quarter of the original size to generate the low-resolution images for deep convolutional neural network training purpose.
Results: train the network with transfer learning and compare the performance with traditional image processing methods (bicubic) and advanced normalization tools (ANTS, R). Through our proposed method, there is a significant improvement for image quality for transferred fMRI than directly resampled or normalized images. Our method also improves the structural similarity index to 94.1%.
Conclusion: experimental results indicate that using deep convolutional neural network can effectively improve multimodal MRI image quality while keeping the motion artifact within a reasonable range. Thus, our method provides a practical solution for multimodal MR image super-resolution and serve as a potential enhancement tool for non-invasive neuroimaging.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email