Transfer learning recently became a popular technique for training machine learning algorithms. A recent development is that it is possible to transfer information from non-medical datasets – for example, the source dataset can be a collection of cat pictures, and the target dataset can be a set of chest CT scans. There is still no definitive answer on what type of data is better to use, therefore more systematic comparisons are needed.
In the group we are leading such a project, where the ultimate goal is to be able to provide advice on what considerations should be made, when choosing a source dataset. There are a number of possible subprojects, which all involve training neural networks on various non-medical and medical datasets. Available subprojects include but are not limited to:
- Interpolating between non-medical and medical data. The goal would be to create various hybrid datasets, and investigate how data similarity influences the quality of transfer learning.
- Combining multiple source datasets in an ensemble of classifiers. When multiple datasets are available, there are different choices on how they can be combined.
Contact: Dr. Veronika Cheplygina (v.cheplygina at tue.nl)
Cheplygina, V. (2019). Cats or CAT scans: transfer learning from natural or medical image source datasets?. Current Opinion in Biomedical Engineering.