post-doc: Cartesian Genetic Programming for 3D biomedical image analysis
Where: Toulouse, 
France
What:
In the past few years, we have used and developed a Cartesian Genetic Programming (CGP) based approach to segment biomedical images. This approach has been showed to be efficient to learn from very few images while keeping high level of interpretability of generated pipelines. Recently, we have proposed the Multimodal Adaptive Graph Evolution (MAGE) approach, which extends the CGP approach to a multi-chromosome representation in order to treat multimodal data. With MAGE, we were able to classify biomedical images and, together with biologists and medical experts, explain the mechanism of decision taking used by the best models. The benefit of our approach are twofold in comparison to SOTA deep learning approaches: (1) CGP/MAGE requires a limited amount of data to learn while keeping good generalisability capacities; (2) CGP/MAGE generates interpretable data analysis pipelines. However, the generated pipelines are still suffering of few points of under-performance in comparison to SOTA deep learning approaches.
The aim of this postdoc is to continue to develop this approach by extending the function libraries to be able to target 3D images, in particular in multimodal brain images. Additionally, we would like to explore automatic abstraction mechanisms in order to generate more complex graphs and, hopefully, improve the performance of the generated pipelines while keeping high level of interpretability.
Duration: 18 months
Location: Toulouse, University Toulouse Capitole, IRIT-CNRS
Salary: based on experience
Who: Sylvain Cussat-Blanc, sylvain.cussat-blanc@ut-capitole.fr
When: Until 2025-12-30 23:00
Presented at next GECCO?: yes