This project investigates domain generalization for medical image analysis using de-identified imaging data collected in Sweden and China. Models developed on one healthcare setting often lose accuracy when applied to another because of differences in scanners, acquisition protocols, patient populations and clinical workflows. We will develop and evaluate deep-learning methods that learn robust image representations across centres and countries, with systematic evaluation on held-out sites and countries. The project will compare domain-generalization strategies, quantify cross-domain performance and identify factors associated with poor transportability. The outcome will be reproducible models and evidence on how multicentre data can improve reliable medical AI. Only de-identified research data without direct personal identifiers will be used.