Analyse quantitative histopathologique par IA pour évaluer des modèles précliniques de ciliopathie et leur réponse à un traitement pharmacologique

Ciliopathies are rare genetic disorders caused by mutations affecting cilia, hairlike organellescrucial for detecting extracellular signals and transporting fluids across epithelial surfaces. Thesedisorders progressively impact multiple organs, including the retina and kidneys. Currently,there are no specific treatments for ciliopathies, highlighting the need for novel therapeuticapproaches. Developing treatments involves assessing molecules and quantifying their effects onaffected organs, a process that is both time-consuming and prone to errors due to the complexityof accurately measuring tissue alterations.In partnership with MEDETIA, a biotech company focused on therapies for ciliopathies, weuse computer-assisted analysis to evaluate disease progression and treatment effects in mousemodels. Employing Digital Pathology and Deep Learning, we've developed models to preciselyquantify histopathologic features of retinal degeneration (in Leber congenital amaurosis models,LCA) and renal malformations (in nephronophthisis, NPHP).In the case of retinal degeneration, we combined semi-supervised segmentation with topolog-ical penalization to assess retinal atrophy in H&E-stained histological sections of LCA modelsdue to CEP290 mutations (LCA10) that causes severe vision impairment from an early age,correlating with reduced retina thickness. This semi-supervised approach alleviates the issue ofannotation scarcity, which is particularly critical in rare diseases. The method is tested on micefor four layers, obtaining a high correlation with manual measurements collected by biologists.Our method offers a promising step forward for quantifying retinal degeneration, with broaderpotential applications in histological studies and ocular disease models.For studies in kidney, we develop a method to automatically detect and measure dilatedtubules. This is relevant as the histopathology of NPHP is characterized by the developmentof cysts within the renal tubules, altering their size. Tubule size is thus a marker to score thedisease and quantify the effect of treatments. A deep learning model segments the cortex, and asecond delineates the tubules in histological sections of mice kidney. Dilated tubules are selectedbased on the geometric properties of their lumen. Using these segmentations to generate indexesthat quantify the effect of NPHP, we obtain high correlation with manual values.Our proposed models are fast and the comparisons with manual measures show high cor-relation, validating our approach. These AI-driven methods enable comprehensive tracking ofciliopathy progression across a substantial population of mouse models. This thesis streamlinesthe drug effect analysis process and facilitates more extensive pre-clinical research efforts.