We investigate the potential of machine learning methods for effective and relevant representation of medical data. To this end, we investigate the following aspects:
- exploiting the latent space to introduce priors on the data representation, in terms of statistical distribution (variational auto-encoders), data-based constraints (semi- and weakly-supervised learning) or application-based constraints (knowledge from biophysical models);
- relating multiple data descriptors, of high-dimensionality and of heterogeneous types, with potential redundancies, either by combining them (multivariate data fusion, multiple kernel learning) or by transferring knowledge from one representation to another (domain adaptation);
- weighting the confidence in the provided output, by means of probabilistic decisions or propagation of uncertainties during the learning process (Bayesian neural networks).