With promising results of machine learning based models in computer vision,\napplications on medical imaging data have been increasing exponentially.\nHowever, generalizations to complex real-world clinical data is a persistent\nproblem. Deep learning models perform well when trained on standardized\ndatasets from artificial settings, such as clinical trials. However, real-world\ndata is different and translations are yielding varying results. The complexity\nof real-world applications in healthcare could emanate from a mixture of\ndifferent data distributions across multiple device domains alongside the\ninevitable noise sourced from varying image resolutions, human errors, and the\nlack of manual gradings. In addition, healthcare applications not only suffer\nfrom the scarcity of labeled data, but also face limited access to unlabeled\ndata due to HIPAA regulations, patient privacy, ambiguity in data ownership,\nand challenges in collecting data from different sources. These limitations\npose additional challenges to applying deep learning algorithms in healthcare\nand clinical translations. In this paper, we utilize self-supervised\nrepresentation learning methods, formulated effectively in transfer learning\nsettings, to address limited data availability. Our experiments verify the\nimportance of diverse real-world data for generalization to clinical settings.\nWe show that by employing a self-supervised approach with transfer learning on\na multi-domain real-world dataset, we can achieve 16% relative improvement on a\nstandardized dataset over supervised baselines.\n