Interactive learning in medical imaging AI: exploiting domain knowledge provided by on-line user feedback

Deep learning for automated delineation and quantification in medical image analysis is complicated by the fact that expert annotated data are scarce, ground truth delineations are subject to observer variability and validation in the clinical setting is needed to get insight in the clinical requirements. Instead of striving for fully automated solutions, keeping the expert clinical user in the loop allows to interactively collect high-quality feedback on the performance of the AI-tools and the user expectations that helps to improve these tools and that offers new possibilities for including rich domain-specific knowledge in deep learning approaches. By collecting user feedback on-line, novel information becomes available to improve the quality of subsequent predictions (e.g. learning from mistakes) or to provide the user an estimate of the reliability of the predictions (e.g. related to local image ambiguity or global image quality). In this project, the concept of interactive learning will be investigated and explored in various real-world medical image segmentation applications in close collaboration with clinical users to develop proof-of-concepts and integrate these in the clinical workflow.