Past 2026 Schedule
Thursday, April 30, 2026, 14:00, Amphithéâtre Jean-Paul Dom
Speaker. Antoine Bralet, ICube (Université Strasbourg)
Title. From Landslides to Mitochondria: Learning Meaningful Latent Representations for Segmentation
Abstract. This presentation explores different deep learning training strategies for learning meaningful representations for image segmentation. First, the focus is on using explainability to force a deep classifier network to detect slow-moving instabilities on interferograms. Constraining the class activation map (CAM) of the network to fit the segmentation of the instability enables the network to better detect instabilities and localize them precisely on the images. Although it is lighter than standard segmentation networks, it focuses on human-like features to detect moving areas. The second part of the talk focuses on ongoing work demonstrating the usefulness of Visual Language Models (VLMs) for segmenting organelles in biomedical images. Including text in the training process aims to facilitate the adaptation of the model to different biomedical modalities by providing a contextual description of how the images were acquired. However, before addressing domain adaptation issues, the relevance of existing pre-trained VLMs is questionable: (1) Traditional VLMs are trained for classification. Thus, applying their knowledge to a segmentation task is challenging. (2) Biomedical images are not representative of the datasets used during training which may question the “zero-shot” capabilities of VLMs for these modalities. The first experiments applying VLMs to biomedical image segmentation illustrate promising results as well as gaps that remain to be addressed.
Thursday, April 30, 2026, 15:15, Amphithéâtre Jean-Paul Dom
Speaker. Vincent Blot, LISN (CNRS, Paris-Saclay University) & Quantmetry
Title. Prédictions conformes et contrôle du risque dans les modèles de vision par ordinateur pour améliorer la performance et la prise de décision humaine
Abstract. Ce travail porte sur la quantification d'incertitude en vision par ordinateur à l'aide de méthodes de prédiction conforme, offrant des garanties statistiques sans hypothèse sur la distribution des données. Trois applications sont présentées : (1) le contrôle certifié de la précision en détection de follicules ovariens dans des lames histologiques, via le cadre Learn-Then-Test ; (2) une méthode de segmentation adaptative (AA-CRC) qui ajuste les seuils par image grâce à des plongements neuronaux, améliorant la précision à rappel garanti ; (3) l'estimation de champs de température à partir de champs de vitesse en convection turbulente, utilisant une architecture mixture-of-experts (U-Net) avec abstention calibrée par pixel. Le fil directeur est l'obtention de garanties fiables en échantillon fini, de manière post-hoc et sans ré-entraînement des modèles.