Xianqi Li1, Bernhard Strasser1, Kourosh Jafari-Khouzani2, Daniel P Cahill3, Jorg Dietrich4, Tracy T Batchelor4, Martin Bendszus5, Ulf Neuberger6, Philipp Vollmuth6, and Ovidiu Andronesi7
1Radiology, Massachusetts General Hospital, Charlestown, MA, United States, 2IBM Watson Health, Boston, MA, United States, 3Neurosurgery, Massachusetts General Hospital, Boston, MA, United States, 4Massachusetts General Hospital, Boston, MA, United States, 5Heidelberg University Hospital, Boston, MA, United States, 6Heidelberg University Hospital, Heidelberg, Germany, 7Massachusetts General Hospital,, Charlestown, MA, United States
1Radiology, Massachusetts General Hospital, Charlestown, MA, United States, 2IBM Watson Health, Boston, MA, United States, 3Neurosurgery, Massachusetts General Hospital, Boston, MA, United States, 4Massachusetts General Hospital, Boston, MA, United States, 5Heidelberg University Hospital, Boston, MA, United States, 6Heidelberg University Hospital, Heidelberg, Germany, 7Massachusetts General Hospital,, Charlestown, MA, United States
We developed deep learning super-resolution MR spectroscopic imaging
(MRSI) to map tumor metabolism in patients with mutant IDH glioma. The preliminary results on simulated and in vivo data indicate that the
proposed method is effective in enhancing the spatial resolution
of metabolite maps.
Fig.5: Upsampled HGG metabolic maps with size 184 x 184
using three measured patients data. From left to right (top), FLAIR,
low-resolution (LR) HGG with size 46 x 46 after spectral quality (SQ) control, missing data
filling in by inpainting (IPT) method and nonlocal means denoising (NLMD),
super-resolution (SR) maps by Bicubic, weighted total variation (wTV), Unet and
GAN. From left to right (bottom), LR HGG with size 46 x 46, LR HGG with size 46 x 46 after SQ control, missing data filling in by IPT
method , SR maps by Bicubic+FNLM, wTV+FNLM, Unet+FNLM and GAN+FNLM.
Fig.4: Upsampled total N-acetylaspartate
(tNAA) maps with size 184x184
using three measured healthy subjects data. From left to right (top),
high-resolution (HR) MEMPRAGE (MPRG), low-resolution (LR) tNAA with size 46 x 46 after spectral quality (SQ) control, missing data
filling in by inpainting (IPT) method and nonlocal means denoising (NLMD),
super-resolution (SR) maps by shown methods. From left to right (bottom), LR tNAA with size 46 x 46, LR tNAA with size 46 x 46 after SQ control, missing data filling in by IPT
method , SR maps by shown methods.