Enhancing Brain MRI Super-Resolution: A Comparative Study of DCNN, GAN, and Vision Transformer Models

Abstract

This study investigates the performance of three deep learning models-deep convolutional neural network (DCNN), generative adversarial network (GAN), and vision transformer (ViT)-for super-resolution (SR) of brain magnetic resonance imaging (MRI) images. The models were evaluated using mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) to assess both pixelwise accuracy and perceptual quality. Results show that the DCNN model outperformed the others in PSNR, achieving 21.24 dB, indicating better pixel-wise accuracy. The ViT model achieved the highest SSIM of 0.6424, reflecting the best structural similarity. Meanwhile, the GAN model showed the best LPIPS score of 0.2133, demonstrating superior perceptual quality. These findings suggest that each model excels in different aspects, highlighting the potential for optimizing MRI resolution in clinical workflows by selecting the most suitable model for specific needs. © 2025 IEEE.