DENOISING OF 7 TESLA MAGNETIC RESONANCE SPECTROSCOPIC IMAGING DATA USING SELF-SUPERVISED LEARNING

Abstract Magnetic Resonance Spectroscopic Imaging (MRSI) extends the capabilities of conventional magnetic resonance imaging (MRI) by providing voxel-wise metabolic information that reflects the biochemical state of tissues. Its ability to measure metabolite distributions non-invasively makes it valuable in diagnosis and monitoring of neurological disorders, cancer, and metabolic diseases. Despite these advantages, the practical use of MRSI is limited by its susceptibility noise arising from hardware constraints, patient motion, magnetic field inhomogeneities, and inherently low metabolite concentrations. This noise obscures spectral peaks, reduces quantification accuracy, and complicates clinical interpretation. Classical denoising methods such as Fourier filtering, wavelet transforms, low rank approximation and singular value decomposition can suppress noise but often risk altering key spectral features and require manual tuning. More recently, deep learning approaches such as convolutional neural networks, autoencoders, and generative adversarial networks have shown strong performance in MRI and 1D MRSI enhancement, yet they still demand a ground truth dataset for training, creating high computational costs and the risk of overfitting. To address these limitations, this thesis introduces a self-supervised denoising framework for 5D MRSI data obtained from six subjects using 7T Scanner at High Field MR Center, Medical University of Vienna. The method combines a 3D U-Net trained with a Noise2Void strategy to denoise spatial components, and an LSTM trained using Noise2Noise approach to learn correlations along the spectral dimensions. Each component data was separately denoised using Leave-out-one cross validation (LOOCV) and concatenated to reconstruct back to original size The results show substantial improvement in both spatial and spectral components. In the spatial domain, the model achieved a peak signal-to-noise ratio (PSNR) of 32.71, which is 53.9% increase from low rank denoised data. The structural similarity index (SSIM) reached 0.8937, and the correlation coefficient was 0.9842, reflecting near-perfect alignment with the underlying signal patterns. In the spectral domain, the denoised output achieved a PSNR of 30.26, SSIM of 0.7945, and correlation of 0.9185, demonstrating that spectral peaks and baseline features were well maintained. These confirm that the proposed model can recover meaningful metabolic information from highly noisy inputs without ground truth data. This advancement is clinically significant because it can be used to enhance metabolite quantification, enable reliable diagnostic interpretation, and support broader clinical use of MRSI in neurological, oncological, and metabolic disease assessment.