The use of diffusion models in the reconstruction of accelerated MRI acquisitions

Miguel Ferreira Cardoso do Espírito Santo Amil

Key information

Authors:

Miguel Ferreira Cardoso do Espírito Santo Amil (Miguel Ferreira Cardoso do Espírito Santo Amil)

Supervisors:

Rita Homem de Gouveia Costanzo Nunes (Rita Homem de Gouveia Costanzo Nunes); Teresa Matias Correia

Published in

December 3, 2024

Abstract

Magnetic Resonance Imaging (MRI) is a widely used imaging modality, particularly for soft tissue imaging. However, its long acquisition times can limit its use, especially in scenarios where patient tolerance and ability to cooperate might be reduced. MRI acquisition can be accelerated by undersampling, however, such techniques introduce noise and artefacts, degrading image quality. Conventional acceleration approaches, such as Compressed Sensing, have been used to address this problem and achieve moderate acceleration factors. Recently, deep learning-based reconstruction methods have emerged as an alternative to further accelerate acquisitions. This thesis proposes a reconstruction approach that combines Compressed Sensing with Deep Learning, more specifically by incorporating a Denoising Diffusion Probabilistic Model (DDPM) to enhance reconstruction quality. The network was based on MONAI’s open-source framework the model was fine-tuned using OPTUNA, a hyperparameter optimisation framework. We aim to use DDPMs to remove undersampling artefacts, substituting the traditional usage of wavelets for denoising in Compress Sensing, followed by a data consistency term, guaranteeing that the reconstruction matches the sampled data. Even when using a simple network, the obtained results are promising. For acceleration factors beyond 2-fold, our proposed approach outperforms typical Compressed Sensing when evaluating quantitatively the reconstructed images in terms of Mean Squared Error (MSE) and Multi-Scale Structural Similarity Index (MS-SSIM). Thus our approach is simpler, more versatile and accessible to a broader range of users when compared to state-of-the-art networks, which use more complex approaches such as self-supervised learning and adapting priors. Moreover, the proposed methods outperform traditional and compressed sensing approaches.