Project 13: Visualisation of Deep Learning on Biomedical Data for Improved Interpretability

Christopher Heje Grønbech

Advisors: Magnus Fontes & Carl-Johan Ivarsson

Project description

Deep learning has been successful over the last decade in computer vision mainly due to highly increased computing power coupled with growing and massive sample sets of images. In order to reach comparable success in life science, we need innovative definitions of what constitute the underlying sample sets. These choices of definitions go hand in hand with the following interpretation of extracted features from the different layers in the deep machine learning architectures. Novel and information-dense visualisations of extracted features and the correlation and relations between features as well as between extracted features and underlying variables defining the samples will be crucial for the ultimate biomedical interpretations and analyses.

As part of this project, I will develop novel visualisation methods and connected machine deep learning algorithms for the interpretation and integrative analysis of high-throughput biomedical data.
These new methods will be applied to relevant biomedical data such immunological patient data, and they will be implemented as software packages for integration with Qlucore’s software platform.

The methods and algorithms will primarily focus on deep generative modelling for representation learning, clustering, and prediction of biomedical data.
In addition to modelling different kinds of data separately, it is also the intent to jointly model several data sources in one end-to-end method to improve the interpretability of the learnt representations.
Although inference is fast, that is finding representations of new data in a trained model, training these models on high-dimensional biomedical data is more time-consuming than linear methods like principal component analysis and factor analysis.
Therefore, methods like transfer learning and one-shot learning to reuse models and speed up training will also be investigated.


  1. Visualising patient state models for medical decision support
    Host: ETH Zürich
    Planned date: February 2021 – April 2021
  2. Visualising biobank-scale health data
    Host: University of Tartu
    Planned date: March 2022 – May 2022