You are here : EUR MANUTECH-SLEIGHT > EUR MANUTECH SLEIGHT > SLEIGHT's research projects

XDeepCell - Research project

Explainable deep models in cell imaging: application to the analysis of structural changes in human cells for diagnostic purpose

PhD student: Martin BLANCHARD, ED 488 SIS (Science, Engineering, Health)

ABSTRACT

This project jointly addresses a biological morphological-characterization problem and a computer vision problem based on advanced deep learning methods. On the biology side, the goal is to understand structural modifications of cells after treatment and to propose a statistical, quantitative method for characterizing these changes. On the computer vision side, the goal is to design explainable representation-learning models suited to poorly annotated data. We first built a dataset of endothelial cell images acquired by immunofluorescence microscopy. These blood-brain-barrier cells are organized into a reference population and several altered populations exposed to an inflammatory agent. We then developed several self-explainable neural architectures based on the concept of prototypes, capable of automatically learning the characteristic structures distinguishing a reference population from an altered one. In particular, we introduce a new approach, ProtoGMVAE, built on a variational autoencoder with a Gaussian mixture prior.


 



PUBLICATIONS

  • M. Blanchard, O. Delézay, C. Ducottet, D. Muselet, Delving into the Explainability of Prototype-Based CNN for Biological Cell Analysis, Proceeding of 2024 IEEE International Conference on Image Processing (ICIP), IEEE, Abu Dhabi, France, 2024: pp. 2909–2915. 10.1109/ICIP51287.2024.10647331
  • M. Blanchard, C. Ducottet, D. Muselet, O. Delézay, ProtoGMVAE: A Variational Auto-Encoder with True Gaussian Mixture Prior for Prototypical-based Self-Explainability, Proceedings of the 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), open access



CONFERENCES

  • M. Blanchard, O. Delézay, C. Ducottet, D. Muselet, Delving into the Explainability of Prototype-Based CNN for Biological Cell Analysis, 2024 IEEE International Conference on Image Processing (ICIP), Abu Dhabi, United Arab Emirates
  • M. Blanchard, C. Ducottet, D. Muselet, O. Delézay, ProtoGMVAE: A Variational Auto-Encoder with True Gaussian Mixture Prior for Prototypical-based Self-Explainability, 2026 IEEE/CVF Winter Conference on Applications of Computer Vision, Tucson, AZ, USA