Machine learning has become an essential tool for extracting meaningful information from complex and noisy data. In this seminar, research on robust machine learning methods originally developed in the context of biomedical signals and imaging will be presented, together with perspectives on how these approaches can be transferred to tomography at synchrotron facilities.

Examples will be given on how challenges such as noise, incompleteness, and real-time constraints arise in biomedical applications, and how methods such as self-supervised denoising, hybrid physics-informed models, and adaptive learning strategies can address them. Analogies with tomography will then be discussed, in particular in the context of the Extremely Brilliant Source (EBS) upgrade at the ESRF, which has led to unprecedented data volumes and the need for online feedback and smarter acquisition strategies.

Finally, perspectives for applying and extending these machine learning approaches to synchrotron tomography will be outlined, with the goal of enabling faster, more robust, and more informative experiments.