Denoising Preclinical MRI with Vendor-neutral Deep Learning-based Image Reconstruction.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Shiga University of Medical Science, Seta-tsukinowa-cho, Otsu, Shiga, Japan, 520-2192. Electronic address: [email protected].
- Department of Radiological Technology, Kyoto University of Medical Science, 1-3 Sonobecho Oyamahigashimachi Imakita, Nantan, Kyoto, Japan, 622-0041.
- Department of Radiology, Shiga University of Medical Science, Seta-tsukinowa-cho, Otsu, Shiga, Japan, 520-2192.
- SwiftMR Research, AIRS Medical Inc., 13-14F, Keungil Tower, 223, Teheran-ro, Gangnam-gu, Seoul, Republic of Korea, 06142.
Abstract
Deep learning-based image reconstruction (DLR) is a widely adopted technology in clinical settings, and recently, the development of vendor-neutral DLR software has been actively pursued. It is valuable to assess the applicability of this technology to MR images in animal experiments. We evaluated the effectiveness of a vendor-neutral DLR method, trained on human clinical MR images, in denoising preclinical MR images of mouse brains, which were not part of the training data. Six FVB mice were scanned using a 4.7T MRI system. Axial 2D T2-weighted images (T2WI) and sagittal 3D T1-weighted images (T1WI) were acquired with varying numbers of signal averages. DLR software was applied at three denoising levels (mild, moderate, and strong). Signal-to-noise ratios (SNR), contrast-to-noise ratios (CNR), and sharpness index were evaluated across the original and DLR-processed images. Increasing the denoising level markedly reduced noise in the original images, making anatomical structures more recognizable. The DLR method effectively reduced noise without compromising anatomical details, as shown by enhanced clarity of fine structures in the processed images. SNR and CNR exhibited a monotonic increase with higher denoising levels, while the sharpness index improved, particularly from original to mild denoising levels. The direct applicability of a vender-neutral DLR method trained on human clinical MR images to mouse brain MR images had not been previously evaluated. The vendor-neutral DLR method substantially denoised preclinical MR images of mouse brains.