0148

View Abstract / View Presentation

Deep Learning Enables A Half Z-spectrum Sampling-based B0 Inhomogeneity Correction for CEST MRI

Yiran Li¹, Danfeng Xie¹, Dushyant Kumar², Abigail Cember², Ravi Prakash Reddy Nanga², Hari Hariharan², John A. Detre³, Ravinder Reddy², and Ze Wang¹
¹Department of Diagnostic Radiology and Nuclear Medicine, University of Maryland School of Medicine, Baltimore, MD, United States, ²Department of Radiology, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, United States, ³Department of Neurology, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, United States

This study presents a DL based framework for correcting B0 inhomogeneity for GluCEST imaging using fewer acquisitions. Based on 3 or 5 positive offset CEST images, the proposed method can save >80% of CEST imaging acquisition time as compared to the conventional protocol.

Fig. 2 The architecture of DL-B0GluCEST-HS is an enhanced deep residual network. The first layer consists of 32 convolutional filters with 3×3 kernel size for each input image. After concatenating them as one channel and going through another convolutional layer, the subsequent layers include 8 consecutive WDSR blocks, which contains 2 convolutional layers and 1 activation layer. Another convolutional layer was attached to the end to get the B0 corrected ±3 ppm image with additional input from the concatenated layer. The subtraction is calculated to obtain B0 corrected CEST image.

Fig. 4 GluCEST ratio maps of a subject were calculated by different methods. Row (a)(c)(e) are GluCEST results and row (b)(d)(f) are differences map between the labeled method and the gold standard conventional approach [10]. The number in the methods indicates how many downfield offset images were used as the input to the DL networks.