Motivation: The image resolution in non-focused planes of 3D MRI volumes is often poorer. Goal(s): This study aims to improve 3D brain MRI resolution across all planes using super-resolution (SR) techniques with a U-Net model. Approach: The images were initially downsampled by scale of 2 and Gaussian blurred which were normalizer and paired with high-resolution images used for training the model. The model's weights were adjusted based on L1 and SmoothL1 loss functions, with learning rate scheduling. Results: The predicted images exhibited significantly improved resolution compared to downsampled images, demonstrating effectiveness of our super-resolution model by reducing the loss from 99 to approximately 5. Impact: Utilizing SR for 3D MRI images is uncommon, yet it significantly enhances MRI efficiency and resolution. The proposed architecture reduces computational costs while improving results, facilitating quicker MRI execution without compromising image quality.
Singh et al. (Tue,) studied this question.