Chronic kidney disease and kidney failure represent a major public health problem. Given the steadily rising number of cases, kidney transplantation is considered the optimal treatment strategy. It offers the advantage of improving survival rates while providing a higher quality of life and lower costs. For kidneys intended for transplantation, it is essential to quickly determine the functional status and the optimal method of kidney preservation, which remains a challenge to this day. Surgeons generally use suboptimal kidneys or exclude potentially better grafts. More broadly, the choice of imaging techniques to analyze the kidney in various clinical settings, including transplantation, is a research topic in its own right. Magnetic resonance imaging (MRI), in particular, shows great potential as a non-invasive method for obtaining structural and functional information. However, the volume and complexity of the data it generates remain a major barrier to its full exploitation. Machine learning, in general, and deep learning, in particular, are widely studied scientific fields found in numerous applications and research areas. Learning-based methods enable a computer to construct complex concepts from simpler ones. Recent advances in medical imaging and machine learning have prompted many researchers to explore the idea of augmented anatomical and functional imaging for diagnostic assistance. By augmented imaging, we mean artificial intelligence (AI) models designed to assist radiologists and enable them to make optimal diagnoses. Our thesis work aims to improve the assessment of kidney graft quality using MRI and machine learning techniques. This work comprises three applications belonging to two main tasks: super-resolution and ultra-high-field MRI synthesis for image quality enhancement; and cross-modal translation. Note that for practical reasons explained in the document, a significant portion of our work was conducted on human brain data. In the first application, we develop a method based on self-supervised models to solve the super-resolution problem on routine 3T MRI using training on both paired and unpaired data. Evaluation of our results shows that the proposed methods can produce high-resolution output from low-resolution input with low distortion. Furthermore, the solution we explore overcomes the limitation of existing methods that require aligned sample pairs. In the second part, we aim to synthesize ultra-high-field MRI data (7 Tesla, or 7T) from 3T volumes. The proposed model achieves convincing results on both objective and subjective metrics. The final models can operate stably on 3D brain MRI, which is very promising. In our latest work, we focus on the task of multimodal translation. The models are designed to generate high-precision volumes across different modalities such as T1⟷T2, T1 ⟷T1c, or T1⟷T2-Flair. Current work focuses on translation from T1 MRI to its contrast-enhanced version, T1c, as this scenario holds significant potential in light of the precautionary principle regarding gadolinium injections for obtaining T1c sequences. A comparative study between methods from the literature and our methods derived from previous work is presented. The results demonstrate that our methods yield stable results across the research dataset.
Author : Quang Huy Do
Management team :
- D.Helbert
- P.Bourdon
- R.Guillevin
Thesis defended on November 16, 2022
Accepted publications :
Huy Do, Pascal Bourdon, David Helbert, Mathieu Naudin, Rémy Guillevin, « 7T MRI super-resolution with Generative Adversarial Network« , IS&T Electronic Imaging 2021 Symposium, 2021, San Francisco, United States, conférenc en ligne