📄 Full Dissertation Report
🎬 Watch the Project Presentation

🌟 Introduction & Project Overview

Medical imaging, particularly Magnetic Resonance Imaging (MRI), is crucial in clinical diagnosis due to its detailed visualization of soft tissues. However, high-resolution (HR) MRI scans require extended scanning times, potentially causing patient discomfort and increased operational costs. To overcome this, researchers are exploring image super-resolution (SR) techniques that transform low-resolution (LR) MRI images into high-resolution images, maintaining critical diagnostic information.

This dissertation investigates two innovative methods to address MRI image super-resolution: Low-Rank Matrix Completion (LRMC) and Deep Learning (DL). The project rigorously compares these methods on the IXI dataset, focusing on T2-weighted MRI scans—renowned for their excellent soft tissue contrast.

🎯 Low-Rank Matrix Completion Approach (ELRTV)

The LRMC technique leverages the inherent low-rank structure of MRI images, enhancing image resolution by reconstructing missing or corrupted image data. The project introduced an Enhanced Low-Rank Total Variation (ELRTV) algorithm, specifically optimized for MRI, featuring:

High-frequency emphasis for sharper edge reconstruction.
Adaptive non-local means regularization for intelligent noise handling.
Singular value thresholding for robust low-rank approximation.
Iterative back-projection to enhance reconstruction fidelity.

This method successfully enhanced image quality at moderate scaling factors (2x SR), but struggled with extreme downsampling (4x), highlighting its limits in generating fine anatomical details without adequate initial data.

🚨 Remarkable Results with Deep Learning

The Deep Learning approach—particularly the Local Implicit Image Function (LIIF) combined with Enhanced Deep Residual Networks (EDSR)—showed exceptional results:

2x Upscaling: Achieved remarkable PSNR and SSIM improvements (+8.59 dB and +0.0557), outperforming conventional and LRMC methods.
4x Upscaling: Delivered significant improvements (+4.10 dB and +0.0798), maintaining critical medical image details at higher magnifications.
Off-domain Robustness: Successfully applied trained models to scaling scenarios beyond their initial training conditions, showcasing flexibility vital for diverse medical applications.

These DL models incorporated advanced mechanisms like Spatial Multi-scale Attention (SMAM) and Edge Enhancement Modules (EEM), effectively preserving subtle anatomical details, textures, and edges crucial for clinical diagnostics.

🖥️ Clinical and Technical Impact

This study clearly highlights DL’s superiority over traditional LRMC in MRI SR, demonstrating both higher image fidelity and greater clinical relevance. The DL methods not only provided visually superior results but also indicated significant potential for enhancing diagnostic accuracy and efficiency in real-world clinical settings.

🔮 Future Directions

Moving forward, this project sets the stage for further research into:

Enhanced DL models capable of handling even larger scaling factors.
Integration of multi-modal medical imaging data to expand the clinical utility of SR.
Optimizing algorithms for resource-limited healthcare environments.