3IA PhD/Postdoc Seminar #30

Flash presentation

A Single Model Strategy for Multi-Domain Cerebrovascular Segmentation

Abstract:
Our work introduces a semi-supervised domain adaptation framework for cerebrovascular image segmentation across diverse modalities. Overcoming limitations of single-modality-focused methods, our approach utilizes annotated angiographies and a diverse range of target datasets from various medical centers, image modalities, and vessel types. Operating within a disentangled latent space, we achieve image-level adaptation without heavy reliance on adversarial training, resulting in a stable and efficient model.

Extended for robust segmentation, the method eliminates the need for ad hoc preprocessing and manual alignment, showcasing its versatility when adapting to large domain gaps. In particular, the framework accomplishes to independently manipulate domain-specific features while preserving crucial spatial information. The evaluation not only demonstrates state-of-the-art performance in the source domain, but also highlights the method's promising potential for accurate cerebrovascular image segmentation across different scenarios. Ablation studies underscore the robustness and efficacy of our approach, emphasizing its potential for accurate cerebrovascular image segmentation across multiple domains.

Physics-inspired learning for fluctuation-based super-resolution microscopy

Abstract: 
Recent approaches to super-resolution fluorescence microscopy exploit the analysis of fluorescence fluctuations over short acquisition times. Mathematically, the problem can be cast in the form of an ill-posed inverse problem and solved using advanced regularisation models and non-convex optimisation tools which are prone to tedious parameter tuning, numerical bottlenecks and potential appearance of artefacts. To overcome these issues, we consider physics-inspired learning approaches to incorporate data knowledge so as to better model data and/or regularisation terms. In particular I will discuss how theoretically-grounded plug & play proximal denoisers can be effectively used to learn implicitly the unknown image prior and present an unsupervised physics-inspired generative adversarial framework reconstructing a high-resolution images temporal sequences of low-resolution data.

Open discussion on the two contributions