Deep learning models have become the method of choice for the automatic detection and segmentation of structures in medical images. Supervised learning methods require training datasets with radiologist's annotations. Supervised learning methods yield excellent results but rely on large annotated datasets to train them, which is expensive and time-consuming, as it requires expert annotators. A variety of methods, including few-shot learning, semi-supervised learning and active learning have been proposed. However, none of them yields satisfactory results, especially for small structures such as lesions and tumors.

In this talk, we present an annotation workflow whose aim is to accelerate the creation of expert-validated annotations while optimizing the annotation and correction time. Specifically, we present two new label-efficient methods. The first is a few-shot learning optimized for small lesions in CT and MRI scans. The second is a semi-supervised method combines the precision of expert annotations with the quantity advantages of pseudo-labeled data. It uses an ensemble of 3D nnU-Net models trained on a few expert-annotated scans to generate pseudo-labels on a large dataset of unannotated scans. We demonstrate both methods on clinically relevant tasks.

During the initial stages of life, the human brain undergoes rapid tissue growth and development after birth. Accurately capturing and describing structural changes from magnetic resonance images (MRI) during this vital period is crucial for gaining new perspectives on healthy brain development and enabling the early identification of neurodevelopmental disorders. While validated brain post-processing and analysis tools exist for adult brains, efforts to directly translate existing adult brain algorithms to pediatric MRI have generally failed, partly due to the poor gray/white matter differentiation of the developing brain and its rapid growth throughout the first years of life. These challenges impair the ability of existing algorithms to accurately segment pediatric brain structures even with high field (1.5T or 3T) MRI systems. In low and middle-income countries, high field MRI systems are rare, due to the cost and maintenance required. More accessible MRI systems like the 0.064T Hyperfine scanner are cheaper and portable, and could help to assess gross anatomical abnormalities; however, image processing challenges are further compounded in these lower resolution imaging environments. We have been developing a set of deep learning based tools for image quality improvement, quality control, and segmentation for ultra low-field MRI. Additionally, we have been determining the useability of ultra-low field MRI in pediatric emergency medicine.