NOISE REMOVAL IN MAGNETIC RESONANCE IMAGING USING 3D DEEP LEARNING MODEL

Magnetic Resonance Imaging (MRI) is a widely used imaging technique for examining brain tissues and diagnosing various conditions. However, MRI images often contain noise caused by factors such as equipment limitations, environmental conditions, patient movement, and magnetic field interference. This noise can obscure critical details, making accurate diagnosis and treatment planning challenging. In this study, the focus is on the removal of Rician noise from MRI images. To address this challenge, two 3D autoencoder models, named M-UNet+ResNet and M-UNet+DenseNet, were developed. These models are based on an enhanced UNet architecture that integrates dense and residual connections, aimed at improving noise reduction capabilities. The models were trained using T1 and T2-weighted MRI images from the IXI dataset, incorporating noise levels varying from 3% to 15%. Their performance was evaluated using metrics such as peak signal-to-noise ratio, structural similarity index measure, and mean absolute error. The results demonstrated that both models effectively reduced noise across various levels, with M-UNet+ResNet generally outperforming M-UNet+DenseNet. Notably, M-UNet+ResNet achieved PSNR values of 38.72 dB and 37.04 dB, and SSIM values of 0.82 and 0.81 in the IXI-HH-T2 and IXI-Guys-T2 datasets, respectively, indicating its strong capability in preserving image quality. This study concludes that incorporating residual connections in DL models enhances their ability to remove noise from MRI images, offering a solution for maintaining the integrity of medical images in clinical settings.