Methods in artificial intelligence guided by biophysics for medical imaging

Background to this project

The research activities of the CREATIS laboratory fall within the field of health technologies, and aim to contribute to predictive and personalized medicine through imaging. Interdisciplinary research brings together experts in image processing and analysis, computer science, physics, instrumentation and radiology. Pathologies such as ischemic heart disease, multiple sclerosis, cancer and stroke are among the areas addressed at CREATIS. The post-doc will be involved in one of the laboratory's cross-disciplinary projects, with the intention of applying for a junior research position within the laboratory.

Objectives

Over the last decade, imaging methods based on deep learning have established themselves as a leading tool in medical image processing. Although they have shown impressive success in various computer vision tasks, their application in medical imaging requires additional considerations due to the need for increased control and the limitations associated with reduced data availability, particularly in 3D imaging. The aim of this post-doc is to focus on new artificial intelligence approaches which, while exploiting deep self-differentiation frameworks, will also incorporate additional modules to explicitly take into account certain physical properties, constraints, or modality knowledge in the network structure or training process.

Methodological innovations

The post-doc will investigate methodological innovations among the following concepts:
- Deep unrolling" methods iteratively apply a deep neural network to simulate the steps of optimization algorithms, enabling an approximate solution to complex problems such as reconstruction, deblurring, super-resolution, segmentation or quantization.
- Deep image prior" methods use the structure and a priori learned by a deep neural network to generate high-quality image reconstructions from incomplete or corrupted input data. The network architecture itself serves as a priori, enabling the generation of plausible images without the need for extensive training data. Modality-specific considerations (CT, MRI, ultrasound, nuclear) as well as biophysical/biochemical modeling of biological tissues will be incorporated into the network design to improve performance and reliability.
- Informed neural networks (PINNs) calculate the derivative of estimated outputs to update additional loss functions that correspond to physical laws or a priori assumptions. In this way, the network is able to optimize (or learn) a solution that must respect the underlying physics.

This methodological research will be applied to one of the following interdisciplinary laboratory projects:
- Multiple sclerosis and neuroinflammation: from preclinical to clinical investigations (MUSIC)
- Radiomics for tumor characterization and treatment response (TUMOR-ID)
- Optical tissue imaging (TipTop)
- Multimodal multiparametric imaging of musculoskeletal and myocardial lesions (IDM4)
- Functional imaging and modeling of the lung (FILM)

Skills

PhD in computer science, physics or related fields, with expertise in deep learning techniques and solid knowledge of medical imaging modalities.
- Technical skills: Python, Java, C++, AI libraries (Tensorflow, PyTorch)
- Written and oral synthesis skills
- Language skills: English (reading, writing, speaking), French preferred
- Writing skills (reports, publications)
- Ability to work in a team/collaborative environment
- Autonomy, organizational and reporting skills

Contacts

If you would like more information about this post-doc position, please send an e-mail to the following address: EqDir@creatis.insa-lyon.fr

If you are interested in this post-doc position, you must register on the CNRS website by clicking here.