French
English
Lyon / San Franciso collaboration
The PhD will be carried out partly at the CREATIS laboratory in Lyon www.creatis.insa-lyon.fr and partly at Yan Li's laboratory at the University of California San Fancisco https://lilabimaging.ucsf.edu as part of an international thesis co-direction.
Contexte:
The analysis of magnetic resonance spectroscopy (MRS) signals has demonstrated its potential as an approach to understanding the major metabolic and neuronal pathways involved in brain disease. However, the analysis of spectra acquired in a clinical context, with current methods, is confronted with a lack of robustness, a lack of objectification of results, and automation of processing processes which reduces the potential of MRS in the clinic, and which deep learning approaches could resolve. Indeed, current methods are based on parametric model fitting. They can be prone to modeling bias, with results depending on the quantification method used, or even the parameterizations used for the same method. Deep learning, on the other hand, will extract maximum information from databases of real spectra, and offer analysis that is not subject to the simplifications of modeling. However, the impossibility to get the ground truth from real invivo MRS acquisition prenset the use of supervised deep learning.
State of the art:
In the literature, machine learning has been explored and used for MRS data analysis, mainly for spectral classification (tumor and non-tumor spectra) and quality control. However, unlike medical image analysis, the analysis of in vivo MRS signals has only just begun to benefit from recent advances in deep learning [1], [2] [3], with problems very different from those encountered in high-resolution NMR used in chemistry or medical imaging.
In [1], a convolutional auto-encoder is trained on simulated data to reconstruct an artifact-free spectrogram from data containing artifacts. In [3], a CNN is trained to predict the evaluation of a spectrum (“good” or “poor” quality). For our part, we have been working on deep learning for data quantization and artifact removal. To our knowledge, we were the first to attempt to quantify MR spectra on synthetic data using a convolutional neural network [4]. As there is no ground truth for in vivo data, we used simulated data with known concentrations for training and testing. The simulations were based on the physical model that is generally used in standard quantification methods, which perform a parametric fit of a model to the acquired data. We also explored the UNET architecture for correcting signal artifacts [5] . However, this neural network architecture does not guarantee the preservation of metabolite content or physical properties. We believe that the constraints of generative networks to preserve certain physical properties as proposed in this project will be more suitable and powerful.
Project description
The project proposes to leverage deep learning for the analysis of NMR spectra acquired in vivo from tumor patients, exploring the latest methodological developments in generative, regression and classification models adapted to the analysis of in vivo magnetic resonance spectra. The project will benefit from a library in PyTorch set up as part of a PNRIA project that supports NMR spectroscopy data, can perform data augmentation and develop typical architectures for classification, regression and adapted data generation. Some parameters (metabolite concentrations, spectral behavior of the macromolecule, etc.) have no “real ground truth”, so a fully supervised approach is not feasible. We will therefore be looking for unsupervised, weakly supervised or semi-supervised learning approaches for this problem.
To avoid the pitfalls of overfitting associated with the sole use of simulated data, unsupervised techniques based on generative score-based models [6], also known as diffusion models, will be developed. We will also seek to develop dedicated hybrid AI methods using simulated data with ground truth, real unannotated data (provided as part of clinical studies conducted by the PSYR2 team or publicly available such as the public database “Big GABA” to start the work) simulated data with the physical model of NMR spectroscopy used in “classical approaches”. This work will build on recent advances in probabilistic scattering models, and propose a new approach to incorporating physical properties/parameters into scattering and inversion processes. More specifically, we will explore a probabilistic conditional diffusion model which, in its inverse process, can incorporate certain targeted properties (e.g. healthy/not healthy, containing water residues or not, containing macromolecule expression or not etc.).
In contrast to the problems encountered in computer vision, we have an (incomplete) mathematical model from physics to model the signal. It will be essential to take this physical model into account when learning generative or analysis models, which may lead us to use domain adaptation methods.
Expected results
The project is part of a drive to develop precision medicine, by improving information from in vivo MRS, the only technique to provide endogenous, non-ionizing, non-invasive information on tissue biochemistry. The techniques developed should 1) make it easier to exploit spectroscopy data on a routine basis, 2) make the quantitative data extracted more robust and reliable than existing methods, and 3) could help detect as yet uncharacterized metabolic or spectral profile modifications, which would provide a better understanding of the mechanisms underlying disease, and aid therapeutic choices.
Application
Background: applied mathematics, machine/deep learning or signal processing
Good software development skills (ideally with python/pytorch)
Taste for working in a highly multidisciplinary environment (deep learning, MRI physics, medical applications)
Applications (CV, transcript, recommendations,...) should be sent to michael.sdika[@]creatis.insa-lyon.fr and helene.ratiney[@]creatis.insa-lyon.fr
Bibliographie
[1] S. P. Kyathanahally, A. Döring, et R. Kreis, « Deep learning approaches for detection and removal of ghosting artifacts in MR spectroscopy », Magn Reson Med, vol. 80, no 3, p. 851 863, sept. 2018, doi: 10.1002/mrm.27096.
[2] S. S. Gurbani et al., « A convolutional neural network to filter artifacts in spectroscopic MRI », Magnetic Resonance in Medicine, vol. 80, no 5, p. 1765 1775, 2018, doi: 10.1002/mrm.27166.
[3] S. Vaziri et al., « Evaluation of deep learning models for quality control of MR spectra », Frontiers In Neuroscience, vol. 17, août 2023, doi: 10.3389/fnins.2023.1219343.
[4] N. Hatami, M. Sdika, et H. Ratiney, « Magnetic Resonance Spectroscopy Quantification Using Deep Learning », Medical Image Computing And Computer Assisted Intervention - MICCAI 2018, PT I. in Lecture Notes in Computer Science, vol. 11070. Springer, Switzerland, p. 467 475. doi: 10.1007/978-3-030-00928-1_53.
[5] N. Hatami, H. Ratiney, et M. Sdika, « MR spectroscopy artifact removal with U-Net convolutional neural network », in 27th Annual meeting of the ISMRM, Montreal, Canada, mai 2019. https://hal.archives-ouvertes.fr/hal-02129946
[6] J. Ho, A. Jain, et P. Abbeel, « Denoising Diffusion Probabilistic Models », in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2020, p. 6840 6851. 6 mars 2023. https://proceedings.neurips.cc/paper/2020/hash/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html