Computer-aided diagnosis (CAD) has become a major research subject in different domains of medical imaging to assist radiologists during their diagnostic task by providing information on the location and characterization (malignancy score) of suspicious regions of interest. The algorithms used to build CAD systems learn a multiclass (mostly binary) decision model in a multidimensional feature space based on training samples from the different classes of interest. Diagnostic performance of such decision support systems is highly impacted by the quality of the training database that should contain a large number of correctly annotated and homogeneous (ie acquired with similar imaging protocols) cases of all classes. Such a condition is not easily met in clinical practice. One way to handle heterogeneous data is to investigate how transfer learning can adapt to this problem.
The purpose of this study is to propose novel methods in the field of domain adaptation for medical imaging based CAD systems and compare their performance with state-of the art methods. The ultimate goal is to boost the diagnostic performance of CAD systems developed at CREATIS by fusing databases originating from different MR scanners.
This study is part of a joint research project between the team ‘Images and Models’ from the CREATIS lab and the team “Data intelligence” from LAHC (Laboratoire Hubert Curien, St Etienne). A first collaborative study between the two teams has shown promising results in the field of prostate cancer mapping based on multiparametric MRI.
During this internship, the successful candidate will work on extending the learning algorithms previously proposed in the collaboration and studying a new formulation based on the optimal transport problem. Please see the attached file for a more detailed description.
- Strong knowledge in at least one of the following fields are required
- Image processing
- Statistical learning (machine learning)
- Applied mathematics
- The available code is written in Matlab and Python but the candidate is free to use any other programming language.
- Motivation and interest for multidisciplinary research.