Jakob Ambsdorf: Medical image analysis with limited labels

The application of Deep Learning to Medical Imaging has largely been enabled by the careful collection of image labels by medical domain experts. In this talk, I briefly review successful applications developed within SONAI, the Danish Fetal Ultrasound project, and show how they are made possible by collaborations between researchers from medical sciences and machine learning researchers. Through this case study, it is evident that the key part of these collaborations lies beyond simply providing labels for machine-learning datasets. However, the acquisition of manual labels is still a time-consuming and often necessary task that can hinder interdisciplinary teams from focusing on more relevant and productive aspects of research.

With this motivation in mind, the main part of this talk focuses on self-supervised pre-training for label efficient learning in medical imaging, and unsupervised methods for solving tasks entirely without manual annotation. I present two of our recent works in detail: Firstly, AMAES is a framework for self-supervised pretraining of 3D brain segmentation models from MRI, and introduces the to-date largest collection of public pretraining datasets compiled in the 🧠BRAINS-45K dataset. We investigate how pretraining in a masked autoencoder framework can improve segmentation performance in the low label regime. Secondly, and returning to fetal ultrasound, I am presenting research on the unsupervised detection of fetal brain anomalies, without requiring segmentation labels and without relying on rare abnormal examples for training.

Jakob Ambsdorf obtained his MSc in Computer Science at the University of Hamburg, where he specialized in machine learning and explainable AI with applications to computer vision and human-robot interaction. He is now a PhD student at the University of Copenhagen and the Pioneer Centre for AI, working on self-supervised learning for fetal ultrasound imaging with a focus on explainability and anomaly detection applications.