Kaisar Kushibar

Over the past 7 years, my research interests have gradually shifted from Robotics to Natural image processing, and then to Medical Image Analysis. I am very interested in the application of Artificial Intelligence in Radiology and Healthcare. During my PhD, I have contributed to the field of translational methods in brain radiology for multiple sclerosis research. My work and findings were in the areas of deep grey matter brain structure segmentation, supervised and unsupervised methods for domain adaptation in deep learning. Currently, my research focuses on validation of machine learning solutions and data for safe utilisation of AI in healthcare. I am working on developing uncertainty estimation methods in deep learning, bias detection, and AI fairness. As part of the EuCanImage project, which primarily focuses on cancer imaging, these topics are crucial for paving the way of AI solutions into real practice. For example, current deep learning methods often fail silently and do not produce any confidence intervals on their decisions. Moreover, bias in datasets often introduce unfairness to the AI solutions. We have seen several examples of such bias in real world: the gender bias in google translator, which favoured male pronouns when translating from gender neutral language to English (He is a doctor vs. She is a doctor). Similar biases in cancer imaging arise depending on geographical, socio-economic, and ethnic backgrounds – e.g., average breast density difference between Asian and European populations. My main research interests include in detection and prevention of such biases and unfairness in deep learning applications in cancer imaging.