MACHINE LEARNING AND DATA MINING

Deep learning, feature extraction, unusual behavior detection, Helmholtz principle, mining textual and unstructured datasets.
The key aim of the project is to apply modern Deep Learning techniques to medical problems.
The initial problem we have looked is to detect cancers in PET/CT scans, the main challenge of this is that we have only rough labels for the data and no accurate ground truth to train a model on.
More recently we have started to look at adversarial attacks on neural networks. This is a very modern topic of research and hasn't been studied yet when applied to medicine.
The idea is that neural networks can be very easy to fool. This is not an issue for simple image classifiers, but as Neural Networks become popular within medical settings this could be a huge problem. Medical fraud (especially in the USA) costs billions of dollars, plus there is the risk of putting patients in danger.