Uncertainty quantification and sampling optimization for deep learning-based medical imaging
Event details
| Date | 29.08.2019 |
| Hour | 10:00 › 12:00 |
| Speaker | Thomas Sanchez |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Dr. François Fleuret
Thesis advisor: Prof. Volkan Cevher
Co-examiner: Prof. Dimitri Van De Ville
Abstract
Fast, reliable and accurate solutions to inverse problems have always been of prime importance to the medical imaging community. In this report, we discuss recent advances in three complementary elements of any imaging pipeline. First, we dive into one of the earliest deep learning-based reconstruction methods for MRI, which achieved unprecedented quality and speed of reconstruction compared to iterative methods. Secondly, we discuss the problem of sampling optimization, where one seeks the best sampling pattern for a given reconstruction algorithm. However, the clinical deployment of any of these methods would require access to both mean and variance estimates, in order to prevent critical diagnoses from relying on uncertain features. A solution to this problem is presented in the last part of this report, where we cover an approach that allows tractable sampling from the posterior distribution of a reconstructed image.
Background papers
A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction, by Schlemper, J. et al.
Learning-Based Compressive MRI, by Gözcü, B. et al.
Deep Bayesian Inversion, by Adler, J. and Öktem, O.
Exam president: Dr. François Fleuret
Thesis advisor: Prof. Volkan Cevher
Co-examiner: Prof. Dimitri Van De Ville
Abstract
Fast, reliable and accurate solutions to inverse problems have always been of prime importance to the medical imaging community. In this report, we discuss recent advances in three complementary elements of any imaging pipeline. First, we dive into one of the earliest deep learning-based reconstruction methods for MRI, which achieved unprecedented quality and speed of reconstruction compared to iterative methods. Secondly, we discuss the problem of sampling optimization, where one seeks the best sampling pattern for a given reconstruction algorithm. However, the clinical deployment of any of these methods would require access to both mean and variance estimates, in order to prevent critical diagnoses from relying on uncertain features. A solution to this problem is presented in the last part of this report, where we cover an approach that allows tractable sampling from the posterior distribution of a reconstructed image.
Background papers
A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction, by Schlemper, J. et al.
Learning-Based Compressive MRI, by Gözcü, B. et al.
Deep Bayesian Inversion, by Adler, J. and Öktem, O.
Practical information
- General public
- Free