Convolutional Neural Network-based Enhancement Method of 3D MRI

Abstract

Magnetic resonance imaging (MRI) employs a magnetic field and radio frequency signal to generate images of the desired body part. However, due to hardware limitations, the resolution in the slice-select direction is lower than the in-plane direction that generates blurring and anisotropic 3D volume. Several resolution enhancement methods have been widely used that improves the resolution of the slice-select plane. In this paper, a deep learning-based resolution enhancement method is developed to form an isotropic 3D MRI volume with improved slice-select direction resolution. In our proposed method, some isotropic 3D image volumes are trained using a deep learning-based resolution enhancement model for generating a residual volume which is utilized to reconstruct MRI volume with the improved slice-select direction resolution. To assess the performance of the proposed algorithm, we employed both quantitative (e.g., peak signal to noise ratio (PSNR) and structural similarity (SSIM) index) and qualitative measurements. Experimental results showed that the 3D MRI volumes produced by the technique have superior quality to volumes reconstructed using other 3D interpolationbased resolution enhancement methods.