Development of new approaches for generating synthetic diffusion gradient directions in diffusion tensor imaging

Magnetic Resonance Imaging (MRI) is a widely used tool in neuroscience and clinical practice, with Diffusion-Weighted Imaging (DWI) offering detailed information about tissue microstructure. Diffusion Tensor Imaging (DTI) models diffusion as a tensor, which allows for quantitative assessment of tissue organization and connectivity. However, DTI requires lengthy acquisitions with numerous gradient directions and amplitudes, limiting its clinical applicability. This thesis examines the application of deep learning to synthetically reconstruct missing diffusion directions, thereby enriching incomplete datasets and reducing acquisition times. A conditional Generative Adversarial Network (cGAN) based on the Pix2Pix framework was implemented and trained on single-shell and multi-shell datasets acquired at the Oxford Centre for Magnetic Resonance (OCMR). Model performance was assessed using image similarity metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), as well as by comparing diffusion tensor-derived maps across selected brain regions. The networks produced realistic synthetic images, and the resulting diffusion metrics closely matched ground truth values, as confirmed by Bland–Altman analysis. The models accurately reconstructed 20% of missing directions in single-shell data and 18% in multi-shell data. These findings indicate that generative Artificial Intelligence methods can accelerate diffusion MRI, increase efficiency, and support broader adoption of advanced diffusion methods in research and clinical contexts.