Description
Date depot: 19 mars 2021
Titre: INTERPRETABLE DEEP LEARNING IN COMPUTATIONAL HISTOPATHOLOGY FOR ALZHEIMER DISEASE PATIENTS’ STRATIFICATION REFINEMENT.
Directeur de thèse:
Daniel RACOCEANU (ICM)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé: Alzheimer’s Disease (AD), the most frequent neurodegenerative disease, is defined by the misfolding and accumulation of Aß peptides and of tau proteins in the brain. Sporadic AD is most commonly present in later life as an amnestic syndrome. However, the clinical presentation of the patients is heterogeneous and different subtypes of the disease have been described, including a rapidly progressive subtype of AD (rAD). Until now, neuropathological assessment of rAD cases was not able to identify specific neuropathological traits for this clinico-pathological entity, despite its unusual fast progression and clinical presentation leading to frequent misdiagnosis as Creutzfeldt-Jakob disease.
Our hypothesis is that rAD brains, as well as other atypical variants of AD, display subtle histological changes that would be undercovered by high-throughput automated microscopic analysis. The topography and morphology of the tau and Aß aggregates, the two main brain lesions characterizing the disease are heterogeneous. Aß accumulation takes the form of focal deposits or diffuse plaques; tau lesions form the so-called neurofibrillary tangles but also present different morphologies in dendrites or axons. We propose to study the topography and morphology of these aggregates to better understand the morphological substratum of AD heterogeneity. To address this question at a large scale, one needs to develop software systems for the automatic segmentation, annotation and quantitation of brain lesions in histo-pathological whole slide images (WSI).
Therefore, the goal of this study is twofold:
1. to develop fully automated, traceable and explainable artificial intelligence (AI) approaches for the histological location and characterization of the tau and Aß aggregates in whole slide brain images, and to deploy it for routine use on the Histomics core facility of the Paris Brain Institute, as on Sorbonne University’s partners’ platforms.
2. to use the previous analytics tool to study to which extent the topography and morphology of the different peptide aggregates present in the brain can be associated with the diversity of symptoms observed in various AD variants.
We propose to design, test and implement an adaptive semi-supervised deep reinforcement learning pipeline, combining methods able to generate high performances (quality and speed), high traceability / explicability / interpretability and facilitate its usability in biomedical research and discovery. Our pathologists have started to acquire and to extensively annotate a unique set of histological images of postmortem brains from the rare form of rpAD and from other identified AD variants.
Our preliminary results suggest that morphological features analysis is eligible for the first level of stratification. We believe that combining these features with topology and semantic-driven image exploration approaches would be able to guide our research toward a refined stratification. Therefore, causal knowledge-based elements, together with semantic-driven WSI exploration will be likely to create a reusable pipeline, able to structure our experience plan, as to justify / explain the generated results. An interpretable approach needs to be designed using emergent traceability and explicability approaches in Deep Learning, in order to facilitate the biomedical interpretability and the advances of multidisciplinary collaborative research.
The tools within this project will contribute to open-source initiatives, and would be therefore available to the scientific community for replicable massive data analysis. This project will therefore contribute to advance the knowledge in AD and push forward the technological development in this area.
Doctorant.e: Jimenez Gabriel