Description
Date depot: 22 avril 2022
Titre: HistoloG-IA: Digital Pathology AI for the follow-up of pre-clinical therapeutic agents in renal ciliopathies
Directeur de thèse:
Jean-Christophe OLIVO-MARIN (Analyse d'images biologiques)
Encadrante :
Vannary MEAS-YEDID (Institut Pasteur (UMR3691))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: This PhD project aims at designing an integrated, computer-assisted analysis to qualify the effect of pharmacological therapeutic agents in genetically deficient mouse models of chronic kidney disease (CKD) in order to discover biomarkers that help to determine who, when and how to treat patients. The analysis will be performed on WSI of histological sections of mouse kidneys. Other data from different modalities such as chemical biomarkers (urine) or data from flow cytometry will be used to refine the prediction of the effects of therapeutic agents. In order to fully exploit these huge amounts of data, the computational pathology should be performed. Supervised deep learning approaches provide excellent results in digital pathology, but need large datasets of annotations, which could be not available in the beginning. The project will benefit from our previous analysis : i) Icytomine to handle gigapixel images (Gonzalez Obando et al. 2019), ii) the large annotated dataset of human kidney (Sicard et al., 2017). However, this should be adapted to the pre-clinical data, with another staining and specific lesions to nephronophthisis disease which have not been annotated. Therefore, the first challenge is to characterize renal lesions with small imperfect annotated dataset. Transfer Learning, weakly supervised learning or self-supervised learning will be explored. The second challenge is to be able to detect mild effect of the therapeutic agents by considering spatial features and integrating other data modality as well as temporal information. The last challenge is to provide a robust and reproducible methodology to face multiple factors of data variability (staining protocols, image acquisition, missing data) to validate the development of potential drugs.
Doctorant.e: Dorra Benjamin