Angular super-resolution in diffusion-weighted magnetic resonance imaging using diffusion probabilistic models

• Ano de defesa: 2026
• Autor(a) principal: Ávila, André Riesco de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: angular super-resolution; diffusion models; diffusion MRI; diminuição do tempo de aquisição; magnetic resonance imaging; modelos probabilísticos de difusão; redes neurais de super-resolução; reduce acquisition time; ressonância magnética; RM de difusão; super-resolução angular; super-resolution neural networks
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/18/18152/tde-10032026-141324/

Magnetic resonance imaging is a technique capable of producing high-quality images of the human body, being the gold standard in medical imaging for many parts of the body, specially for soft tissues like the brain. Also, it does not expose the patient to ionizing radiation like in X-rays and CT-scans. MRI scans, however, are more expensive and have longer acquisition times when compared to the other methods. Leveraging the MRI technology, it is also possible to obtain images that provide information like the diffusion of water molecules in the brain (diffusion MRI), which can be very useful for detecting earlier strokes and making diagnostics of diseases such as migraine. Diffusion MRI, however, needs to acquire images in many diffusion gradient directions, which take a long time to acquire, on the order of several minutes to more than an hour. To reduce this acquisition time, it is possible to reduce the resolution of the image acquired or reduce the number of diffusion gradient directions. On recent years, deep learning techniques have shown great results on improving the quality of diffusion magnetic resonance imaging with less resolution or less diffusion gradient directions. Based on that, this work proposes a novel method of improving the quality of diffusion MRI with less diffusion gradient directions acquired, by performing angular super-resolution through the use of a super-resolution neural network. The super-resolution neural network used in this work is of the class diffusion models, a class of generative neural-networks that in recent years have been showing outstanding results in the super-resolution task in many applications, and also in many other tasks like denoising and segmentation in medical imaging and even more specifically in MRI. The proposed method was able to improve diffusion metrics obtained with few diffusion gradient directions (8 directions- 6 DWIs + 2 b0s). As diffusion tensor imaging (DTI) was used, fractional anisotropy is the metric most sensitive to number of directions, and in this metric it improved both visually and on the evaluation metrics proposed (SSIM increased from 0.7182 to 0.8028, PSNR from 16.19 to 19.52, the difference in the median values decreased from 46.42% to 14.54%, and ratio of voxels with statistical difference in TBSS reduced from 58.17% to 53.77%). However, the approach also incurs high computational cost and, due to potential generative-AI hallucinations, is best suited for applications that rely on summary statistics rather than direct image interpretation.