Medical image analysis is today mainly done through deep learning. Training deep models require large datasets and advanced GPU hardware, especially when 3D models are used, and for foundation models, GANs and diffusion models. Access to large datasets is in medical imaging hindered by ethics and regulations like GDPR. In this project we will therefore explore three approaches to make it easier to train deep models without having a large dataset on a single computer. The first approach involves training our own foundation model for orthopedics, such that other research groups working on orthopedics can start their trainings from our foundation model (instead of starting from foundation models trained on other types of data). The second approach involves training models through so called federated learning (FL), where smaller datasets are available at each node in a federation. In FL no data are shared between nodes, but a global model is trained by instead sharing model weights between each node and a global server. The third approach involves training generative models such as GANs and diffusion models to create realistic synthetic images. These synthetic images can potentially be shared freely, as they do not belong to a specific person.