37 datascEYEnce! Computer Vision News The one word Cristina probably used most in the interview was “limitations”: her research sheds light on the limitations of data and algorithms, as well as the requirements and mistrust coming from different stakeholders. And in order to evade those limitations, someone first needs to identify them, analyse them, find possible solutions, and document them - and this is what Cristina did. Let’s start with the data. Cristina has worked with a variety of medical data and emphasized the importance of understanding the data you are working with. Somewhat, computer vision techniques are similar across medical imaging modalities. However, there remain considerable differences among modalities, for example, when it comes to vulnerability to adversarial attacks (Bortsova, González-Gonzalo, Wetstein et al., 2021). Different imaging modalities have different limitations and this obviously shall influence the choice of the most suitable algorithms and techniques. In order to find a suitable algorithm, we need to figure out what makes an algorithm suitable for whom. The first step is to start the conversation with relevant stakeholders, such as the clinicians (or healthcare providers) who are meant to use the algorithm. Cristina told me that you have to take notes on variables that clinicians think are important, for instance, the severity of a disease or specific phenotype information. The next step is data exploration and figuring out which data to consider in the analysis. It is indispensable to understand demographics and identify biases that might be present in the data. In the development phase, we shall address potential biases by adequately pre-processing the data and/or adding constraints for bias AI pipeline and stakeholder overview from "Trustworthy AI: Closing the gap between development and integration of AI systems in ophthalmic practice"
RkJQdWJsaXNoZXIy NTc3NzU=