Click here to


Are you sure ?

Yes, do it No, cancel

Deep Learning-Augmented MRI for Quantitative Imaging

Y Wu1*, J Du2 , A Hsu1 , D Hristov1 , W Zhao2 , Y Ma2 , M Chan3 , L Xing1 , (1) Stanford Univ, School of Medicine, Stanford, CA, (2) University of California, San Diego, La Jolla, CA, (3) Memorial Sloan-Kettering Cancer Center, Basking Ridge, NJ


(Tuesday, 7/31/2018) 1:15 PM - 1:45 PM

Room: Exhibit Hall | Forum 6

Purpose: In MRI, a high-spatial resolution is only achievable with long scan time, which is costly, inefficient, and subject to intra-scan motion blurring. The problem of long scan times is magnified in quantitative imaging of bone and cartilage, where a series of images are acquired. Here we propose a deep learning based augmentation technique that is capable of providing super-resolution and high-quality MR images from low-resolution images acquired at reduced scanning time.

Methods: A volumetric deep convolutional neural network was trained to provide the mapping from low-resolution images to high-resolution images. The network architecture consisted of two paths, contracting and expanding, each having six hierarchical levels that were connected via global shortcuts. Within each level on a single path, there were several convolutional blocks, where local shortcuts were employed to facilitate residual learning. The network was trained with the Adam optimization method, and the loss function was the mean-squared-error. For the training of the model, 376 images with a resolution of 256x256x36 was acquired using an Ultra-short TE (UTE) MR sequence. Given the fully sampled images, undersampling was simulated in k-space as radially undersampling in-plane and zero filling through-plane. Image augmentation (rotation, flipping) was applied to substantially enlarge the data sets used.

Results: Image quality improvement using the model was demonstrated and proved to be 4 times more efficient than directly acquiring high resolution images. The convergence was fast, taking only 1000 iterations. The super-resolution processing for a 3D image takes one second.

Conclusion: The volumetric deep convolutional neural network with local and global short connections provides an effective way to obtain super-resolution MR images without the need for elongated acquisition time. Further investigation to quantify image quality and errors are underway.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email