Denoising Cardiac MRI Images Using Deep Learning

Magnetic resonance imaging (MRI) is crucial for cardiac diagnostics, but its cost and infrastructure needs limit accessibility in low-resource settings. Low-field MRI offers a more affordable alternative, though its scans are noisier, affecting diagnostic accuracy. This study proposes a denoising approach using a convolutional neural network (CNN) with a U-Net architecture to enhance low-field cardiac MRI scans. Synthetic Gaussian noise at five intensities was added to images from the ACDC dataset, which includes MRI scans from 150 patients. The model was trained to minimize pixel-wise L2 loss between noisy and clean images. Performance was evaluated using RMSE, PSNR, SSIM, and segmentation accuracy from an external cardiac segmentation model. Compared to a total variation (TV) baseline, the CNN achieved more accurate reconstructions and better preserved anatomical detail, improving segmentation outcomes. These results suggest that deep learning can enhance the clinical utility of low-field MRI and support broader, more equitable access to cardiac imaging.