Computer Vision News - June 2023

25 Exploring Image Augmentations for Siamese ... To learn more about this work, including the team’s challenges and Nandita’s work outside of this paper, please watch our video interview. All images are used with permission from: Nandita Bhaskhar - Robust, Data-efficient, and Trustworthy Medical AI - 2023. property by which these methods learn. The paper explores how to leverage the domain-specific properties of medical images to determine suitable augmentations that enhance self- supervision. “ There are a few things that make our paper unique,” Nandita explains. “There are so many different types of augmentations that can be done for medical images. This is the first paper that did a systematic evaluation of various pairs of augmentations over a largescalewithmultipledatasets .Wealso focus specifically on domain specificity . We introduced a set of augmentations with optimal hyperparameters . It’s a recipe for practitioners to start using these techniques for their models. In our evaluations, we showed that not only are these well-performing augmentations, but they generalise well to different hospitals and domains. In our paper, we show that our models can even generalise to unseen diseases . For example, tuberculosiswas never seen before by our model, but now it’s able to understand something is new, and something is tuberculosis.”