ADVANCING MRI SUPER-RESOLUTION: AN INNOVATIVE DEEP LEARNING APPROACH FOR ENHANCED RADIOLOGICAL STRUCTURE SUPER-RESOLUTION

Abstract

This study focuses on the application of deep learning algorithms to super-resolution in medical MRI images to improve radiological structures, which has, in the recent times become extremely important to the diagnostics. The methodology for the purpose involved preprocessing and augmentation MRI images, to improve model generalization. Multiple algorithms which use differing flows to enhance an image, were selected for evaluation. These included: 1) Bicubic Interpolation, 2) SRCNN, 3) EDSR, and 4) ESRGAN. The models were assessed using key performance metrics like Peak Signal-to-Noise-Ratio (PSNR), Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM). ESRGAN was found to have highest PSNR (33.75 dB) and SSIM (0.98), while one of the lowest MSE (82.44). Conclusively, ESRGAN showed a superior and structural integrity and perceptual quality over all other models on medical images. It was demonstrated that ESRGAN is able to better restore fine details and textures than the state of the art, especially for medical imaging where precision is critical. The qualitative visual assessments also confirmed that ESRGAN dominates in the super-resolution image quality, as its reconstructed images matched high resolution and very precisely maintained critical radiological features. As this study concludes, ESRGAN is the most effective for super-resolution of MRI images, and has great potential for improving the diagnostic capabilities in medical imaging.