Unsupervised anomaly detection in neuroimaging : contributions to representation learning and density support estimation in the latent space

Le 11 avril 2024 à 14h,

Amphi Claude Chappe, Département TC (Bâtiment Hedy Lamarr)

Nicolas Pinon soutiendra sa thèse

"Unsupervised anomaly detection in neuroimaging : contributions to representation learning and density support estimation in the latent space"

 Résumé :

This PhD thesis covers the topic of deep unsupervised anomaly detection (UAD) in neuroimaging. This researchis partially groundedon the UAD model that was proposed in [Alaverdyan, MEDIA 2020], whose novelty was to perform the detection step in the latent representation space by adjusting density support model of the normative distribution. This model developed was applied to the detection of subtle epileptogenic zones in multiparametric MRI and evaluated on a private database. As a first part of this PhD, we optimize the architecture and hyperparameter setting of this UAD model, and evaluate its performance on different open datasets, including the non medical MVTec anomaly detection [Pinon, GRETSI 2023], the WMH challenge [Pinon, MIDL 2023], and the Parkinson's Progression Markers Initiative database [Ramirez, Pinon, MLCN 2021][Pinon, ISBI 2023a. This allows comparison with state of the art deep UAD methods, especially with the most common methods based on reconstruction error in the image space. The second main phase of this PhD work is to build on the limits of this model {Alaverdyan, PIEDIA 2020] and propose original methodological contributions to 1) design patient specific models, relaxing the strong constraint to accurately coregister all control subjects and patients [Pinon, MIDL 2023], 2) provide a probabilistic detection frameworkto enable ensemble  learning and probability calibration, 3) fuse the representation learning step and the outlierdetection step, by proposing a novel end-to-enddeep learning model.

Composition du jury :

La soutenance se déroulera en anglais.