Robust and safe machine learning in medical image analysis

Background Machine learning and specifically deep learning and convolution neural networks (CNN) outperforms traditional medical image analysis for almost all application areas. However, clinical adoption has been slow. Lack of clinical adaptation can be attributed to a lack of robustness and safety in the developed systems.

With robustness we mean the ability of producing high quality output despite challenging or variable input data.

With safety we mean that the failure modes of the system are known and understood and can therefore be mitigated.Purpose The purpose of the project is to increase our understanding on how to develop robust, safe and clinically useful machine learning applications.Project Robustness will be increased by data driven selection of training data and data driven image augmentation.

Safety will be increased by that the network will be able to detect that input data is outside the domain on what it was trained.

Finally, we will present a framework for data driven system tuning to increase development speed and a best practise on how to evaluate medical machine learning systems.Significance In summary, the project will advance the research frontier of applied medical image analysis by novel data driven tools to ensure robustness and safety of real-life machine leaning systems in clinical practice.