Master Internship in AI for biological microscopy

Context
The successful candidate will join the OPIS project team under the supervision of team leader Emilie Chouzenoux.
The project will be carried out in collaboration with E. Chouzenoux and J.-C. Pesquet (OPIS), C. Lefort (XLIM, CNRS, Limoges), and Limoges University Hospital.
The internship will last 5 to 6 months, with considerable flexibility regarding the start date. It is primarily intended for students completing their final-year engineering internship or the second year of a Master’s degree (M2), but applications from engineering students taking a gap year are also welcome.
Assignment
CARS (Coherent Anti-Stokes Raman Scattering) microscopy is an advanced nonlinear optical imaging technique that provides label-free vibrational information about biomedical samples. Broadband laser sources known as “supercontinuum” sources enable the exploration of a wide range of vibrational transitions within the sample. Combined with spectral detection, multiplex CARS (M-CARS) yields hyperspectral images (HM-CARS), in which each pixel contains a 1,024-point spectrum reflecting the chemical composition of the sample.
Using this approach, the project partners
https://www.multi-opto-diag.fr/khi-malma-piq/
recently demonstrated that the so-called “silent region” of HM-CARS spectra contains information that can help discriminate between samples. The biomedical potential of this information remains largely unexplored. Current data analysis methods are laborious and face several practical and methodological challenges.
The main objective of this internship is to develop fast, reliable, and automated methods for extracting information relevant to the characterization of muscle diseases from HM-CARS images. The work will combine AI-based approaches with optimization-based methods, notably by formulating spectral unmixing as an inverse problem and developing advanced proximal optimization algorithms to solve it.
Main activities
Understand the principles of HM-CARS imaging and the challenges of data processing for muscle disease characterization.
Explore the dataset and implement suitable preprocessing methods.
Formulate spectral unmixing as an inverse problem and develop solution methods based on proximal optimization.
Develop supervised learning approaches to exploit the spectral and spatial information in the images.
Evaluate and compare the proposed methods in terms of accuracy, robustness, and computational time.
Write scientific reports and present the results.
Participate in scientific meetings with the project partners.
Skills
Proficiency in the Python programming language and the PyTorch or TensorFlow environment is required.
Experience in machine learning / neural networks is strongly recommended.
Candidates must have validated a course in mathematical optimization.