BIMG-22. DEEP LEARNING SUPER-RESOLUTION MR SPECTROSCOPIC IMAGING TO MAP TUMOR METABOLISM IN MUTANT IDH GLIOMA PATIENTS. (25th March 2021)

• Record Type: Journal Article
• Authors: Li, Xianqi; Andronesi, Ovidiu; Strasser, Bernhard; Jafari-Khouzani, Kourosh; Cahill, Daniel; Dietrich, Jorg; Batchelor, Tracy; Bendszus, Martin; Neuberger, Ulf; Vollmuth, Philipp
• Is Part Of: Neuro-oncology advances. Volume 3(2021)Supplement 1
• Page Start: i5
• Page End: i6
• Publication Date: 2021-03-25
• Subjects: 616.99481
• DOI: 10.1093/noajnl/vdab024.021
• Languages: English
• ISSNs: 2632-2498

Abstract

Abstract: Metabolic imaging can map spatially abnormal molecular pathways with higher specificity for cancer compared to anatomical imaging. However, acquiring high resolution metabolic maps similar to anatomical MRI is challenging due to low metabolite concentrations, and alternative approaches that increase resolution by post-acquisition image processing can mitigate this limitation. We developed deep learning super-resolution MR spectroscopic imaging (MRSI) to map tumor metabolism in patients with mutant IDH glioma. We used a generative adversarial network (GAN) architecture comprised of a UNet neural network as the generator network and a discriminator network for adversarial training. For training we simulated a large data set of 9600 images with realistic quality for acquired MRSI to effectively train the deep learning model to upsample by a factor of four. Two types of training were performed: 1) using only the MRSI data, and 2) using MRSI and prior information from anatomical MRI to further enhance structural details. The performance of super-resolution methods was evaluated by peak SNR (PSNR), structure similarity index (SSIM), and feature similarity index (FSIM). After training on simulations, GAN was evaluated on measured MRSI metabolic maps acquired with resolution 5.2×5.2 mm 2 and upsampled to 1.3×1.3 mm 2 . The GAN trained only on MRSI achieved PSNR = 27.94, SSIM = 0.88, FSIM = 0.89. Using prior anatomical MRI improved GAN performance to PSNR = 30.75, SSIM = 0.90, FSIM = 0.92. In the patient measured data, GAN super-resolution metabolic images provided clearer tumor margins and made apparent the tumor metabolic heterogeneity. Compared to conventional image interpolation such as bicubic or total variation, deep learning methods provided sharper edges and less blurring of structural details. Our results indicate that the proposed deep learning method is effective in enhancing the spatial resolution of metabolite maps which may better guide treatment in mutant IDH glioma patients.