APTOR – Reliable Analysis Procedure of Tomography-imaged Objects using self-learned Representations

AI has shown remarkable transformative results but at the cost of preparing data for training, especially preparing labeled data. Recently, self-supervised vision models have emerged that, without any labeled data, can learn what is depicted in an image if the models are presented with an enormous amount of (unlabeled) training data. These models work for 2D, but what about 3D volumes? Can we get the same performance in 3D? In 3D synchrotron imaging, we have also witnessed a mindblowing development. We can now record terabytes of huge images in a single day, revealing structures that were not accessible earlier. However, the data is so large that we cannot measure the structures using today’s analysis tools. So, how about combining the abundance of data from 3D imaging with self-supervised deep learning models – so-called foundation models? This is exactly what we will do in the RAPTOR project, and, with this, we expect to change the future of large-scale 3D image analysis.