10122- NI-9 DEEP-LEARNING-BASED RECONSTRUCTION OF DIFFUSION WEIGHTED IMAGES OF MALIGNANT BRAIN TUMORS. Authors: Yusuke Sakaguchi, Satoshi Takahashi, Naoaki Ishiguro, Yuichi Suzuki, Hirokazu Takami, Masamichi Takahashi, Shunsaku Takayanagi, Mototaka Miyake, Masashi Nomura, Yosuke Kitagawa, Shota Tanaka, Yoshitaka Narita, Nobuhito Saito, Ryuji Hamamoto Published: 2024 ABSTRACT AbstractBACKGROUNDDiffusion Kurtosis Imaging (DKI) is increasingly used to diagnose malignant brain tumors by assessing non-Gaussian diffusion of water molecules, reflecting tissue microstructure. It aids in preoperative molecular classification and infiltration identification. However, its time-consuming process and complex analysis, requiring Diffusion Weighted Images (DWI) at b=0 and at least two distinct non-zero b-values across 15 axes, hinder its clinical adoption. This study aims to simplify this by using deep learning to reconstruct DWI (b=2000) from DWI (b=0, 1000). MATERIALS & METHODSWe analyzed 157 malignant brain tumor cases from the University of Tokyo Hospital (2016-2023) using the ResViT model for reconstructing high-accuracy medical images, performing 5-fold cross-validation on 282,600 DWI sets. The reconstructed images were evaluated with the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR). RESULTSReconstructed images showed SSIM and PSNR values of 0.956±0.007 and 31.46±1.05, respectively. DISCUSSIONThe deep learning-reconstructed DWI (b=2000) images were comparable to prior studies, suggesting potential for reduced DKI imaging time.