Traumatic brain injuries is one of the major cause of death and disability worldwide especially among young people. The initial injury can later develop into a secondary injury that is particularly lethal. The patient’s survival relies on fast and proper anticipation of secondary injuries by medical worker. However, it is a demanding task for radiologist, that could be eased by automatizing it. In this context, Deep Learning algorithms pipeline would be well adapted to efficiently process the high-dimensional CT-scans. In this context, the very first step of such a pipeline is to locate intracranial hemorrhages (ICH) form CT-scans. Because ICH comes in various shapes, sizes and location, a large number of labelled data is required to train robust and trustworthy models. However it is prohibitively expensive and time consuming to acquire many volumetric labels from trained radiologist. As a results, there is a need for label-efficient algorithms that can exploit easily available unlabelled and weakly labelled data. In this work we explore a varieties of methods that harness unlabelled and weakly labelled data to improves the ICH segmentation. Unlike most studies on ICH, our work relies exclusively on publicly available datasets which allow to easily compare performances with future studies. We further explore unsupervised ICH localization through the anomaly detection principle and we propose a novel approach based on image inpainting. This study demonstrates that unlabelled and weakly labelled data can be used to greatly improve the ICH segmentation performances with a scarce amount of annotations. We further show that our anomaly localization methods enables to extract meaningfully hemorrhages in an unsupervised way, and demonstrate its general capabilities on an anomaly localization benchmark, the MV-Tec dataset.