VNetLS: Supervised 3D VNet-Based MRI Enhancement with Loss Optimization

Abstract

Magnetic Resonance Imaging (MRI) is essential for medical diagnosis, but image quality depends heavily on scanner strength. High-field MRI systems such as 7T provide superior resolution and contrast; however, their high cost and limited availability restrict widespread clinical use in under developed areas. This study proposes a deep learning–based framework, VNetLS, to enhance low-field 3T MRI images into high-quality 7T-like outputs. The method employs a structured preprocessing pipeline, including skull stripping, bias field correction, registration, and intensity normalization, to ensure data consistency. A 3D VNet encoder–decode architecture is used to effectively capture volumetric anatomical features. The key contribution is a multi-component loss function that combines Mean Absolute Error (MAE), Structural Similarity Index Measure (SSIM), and gradient loss, improving image sharpness, structural preservation, and edge detail. In addition, several fine-tuning strategies were explored to further optimize model performance, where replacing the optimizer with AdamW demonstrated the most significant improvement in output quality. Experimental results show that the proposed model outperforms baseline approaches, achieving higher PSNR and SSIM values, indicating improved image quality and anatomical fidelity. This framework offers a cost-effective, software-based solution for MRI enhancement, particularly beneficial for resource-limited healthcare settings. Future work includes expanding paired datasets and incorporating multi-modal MRI inputs to further improve robustness and performance.